Authors: Yepeng Shi Shengliang Fang Shunhu Hou Yuhai Li You Fu Qichen Wang
Radio Environment Map (REM) construction is frequently constrained by sparse and unevenly distributed spectrum measurements. While existing completion methods primarily target Power Spectral Density (PSD) data under random missing patterns, the reconstruction of Reference Signal Received Power (RSRP) maps under structured data loss remains underexplored. This study addresses this gap by proposing a fully convolutional densely connected autoencoder(AE) for RSRP map completion. The encoder stacks dense blocks and transition layers, a bottleneck preserves the latent representation, and the decoder restores spatial resolution through transposed convolution. Both global and local skip connections are incorporated to fuse large-scale structure with fine-grained details. A composite loss function supervises observed and missing regions separately, which preserves the fidelity of known measurements while improving inference over unobserved grid points. Experiments on the public DeepREM dataset under random, spatial, and strip-wise missing patterns show that the method achieves the best or comparable completion accuracy in most tested settings, with the most pronounced performance gains over mainstream baselines under the challenging spatial block-missing case.
]]>Authors: Gali Chelouche-Dwek Peter Fonagy
Background: Teachers routinely encounter children whose behaviour reflects emotional distress and dysregulation, yet they have limited opportunities to practise the relational skills required to respond effectively. These challenges are particularly pronounced in Alternative Provision (AP), which serves children who frequently present with histories of trauma, neurodevelopmental differences, and complex emotional and behavioural needs. Mentalization, the capacity to understand behaviour in terms of underlying mental states, is central to effective relational practice in such contexts. Conversational Artificial Intelligence (AI) may offer a scalable means of supporting this form of skills development, but its feasibility as a teacher-training modality remains largely unexplored. Methods: This mixed-methods proof-of-concept feasibility study evaluated MentaClassAI, a novel AI-based training tool in which educators engaged in simulated voice conversations with AI child characters portraying classroom dysregulation and subsequently received individualised, mentalization-informed feedback. Eleven staff members from a single AP school (four teachers and seven teaching assistants) completed a single training session and were allocated to either a psychoeducation video condition (n = 6) or a no-video condition (n = 5). The video condition received a brief introduction to mentalization and epistemic trust prior to engaging with the simulation. Pre- and post-engagement measures included the Reflective Functioning Questionnaire (RFQ-8) and a Teacher Self-Efficacy Scale. Post-engagement measures included an 18-item acceptability questionnaire, a Technology Acceptance Model scale, and open-ended questions analysed using thematic analysis. Results: Acceptability was high, with 84.8% of questionnaire responses falling within the positive range (overall M = 5.60/7). Feedback accuracy (M = 6.55) and clarity (M = 6.36) received the highest ratings. Participants reported higher teacher self-efficacy after the session than before (d = 1.20, p = 0.003), with 10 of 11 participants demonstrating improvement. Self-reported hypomentalizing was lower after the session (d = −0.86, p = 0.017). Between-condition differences (video versus no-video) were not statistically significant. The video condition scored numerically higher on the directional indicators. Qualitative analysis identified five themes: the value of consequence-free rehearsal; the specificity and usefulness of feedback; appreciation of the focus on the child’s emotional experience; limitations in the ecological diversity of AI child characters; and a desire for more naturalistic interaction. Conclusions: These findings provide preliminary support for the feasibility and acceptability of AI-based mentalization practice for AP staff. The principal value of the tool appears to lie not only in the simulation itself but in the quality of the reflective feedback generated. Although based on a small sample, the observed pre–post changes provide an encouraging signal that may justify a controlled trial. The contribution of pre-session psychoeducation to training outcomes remains an important question for future research.
]]>Authors: Laura Acosta García Julen Cestero Portu Ander García Gangoiti Marco Quartulli
Warehouse operations involve dynamic item flows, fluctuating demand, and heterogeneous layouts, making adaptive decision-making essential for efficient storage and order fulfillment. In this context, reinforcement learning (RL) provides a promising approach for learning adaptive warehouse control policies under stochastic environments. However, evaluating RL-based solutions in real warehouse settings is often costly and time-consuming, motivating the need for realistic and reproducible simulation environments. In this paper, we introduce a configurable warehouse simulation environment modeling stochastic item arrivals, order generation, and internal logistics operations across diverse layouts and workload conditions. Based on this environment, we construct a reproducible experimental testbed composed of multiple scenarios ranging from low-load to highly congested settings. The testbed is publicly released to support reproducible research and comparative evaluation within the research community. We formulate the warehouse management problem as a Markov decision process (MDP) and apply a Maskable Proximal Policy Optimization (Maskable PPO) agent to learn adaptive control policies. The RL-based approach is evaluated across the defined scenarios and compared against heuristic baseline strategies. Experimental results show that the proposed solution achieves performance comparable to a strong greedy first-in, first-out (FIFO) heuristic while improving order fulfillment by up to 13.5 percentage points under challenging workload conditions. These results demonstrate the ability of RL to learn robust warehouse control policies that adaptively optimize performance and maintain operational stability across a wide spectrum of distinct scenarios.
]]>Authors: Oleksandra Topal Inna Novalija Joao Pita Costa Dumitru Roman
This study deploys a scalable machine learning pipeline: combining a transformer-based classifier applied to 2.01 million English-language AI-related news headlines (July 2022–July 2024) with large-language-model and human-annotator validation (three annotators, Fleiss’ κ=0.80) on stratified subsamples, to extract six interpretable, bias-linked discourse indicators computed at the AI-domain level: evaluative orientation (valence), loss salience, narrative drift, exposure-adjusted sentiment, cross-source divergence, and novelty-phase framing. Each operationalizes an established cognitive-psychology construct as a computable property of the information environment associated with biased risk–benefit reasoning. Results show systematic variation across domains: technical and methodological areas such as deep learning and natural language processing exhibit gain-salient framing, while safety-critical topics such as deepfakes (loss-to-gain headline ratio = 3.17) and facial recognition show strongly loss-salient profiles. Cross-model validation using an LLM on a stratified sample of 1000 headlines confirms that domain-level indicator rankings are robust to classifier choice (Spearman ρ=0.83; p<0.001), establishing the rank stability of pipeline outputs independently of the specific classification architecture. As a contextual application, domain-level profiles are mapped to European Union AI governance instruments, documenting parallels between discourse patterns and regulatory risk tiers. The framework provides a scalable, reproducible methodology for monitoring evaluative conditions in technology news across domains, sources, and time.
]]>Authors: Genaro Herrera Cano Michal Dyrda Youssef Beshay David Baltrusaitis Mitch Paro Rafael Olivieri-Ortiz Grigoriy Androsov Antonio Medina Luna Michael Baldwin
Traumatic cervical spine fractures (CSFxs) require timely diagnosis due to associated morbidity. Artificial intelligence (AI)-based decision support systems have been proposed to improve imaging workflow efficiency. However, their performance in non-trauma settings remains unclear. This study evaluated the diagnostic performance of the AIDOC decision support system (DSS) for detecting CSFxs in a non-trauma academic community hospital using a retrospective analysis of 1812 cervical spine CT scans, with radiologist interpretation as the reference standard. Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were calculated for fracture detection and heatmap-based localization. The AI system demonstrated a sensitivity of 72.2% and specificity of 98.1%, with an accuracy of 97.9%. In the context of low fracture prevalence (0.99%), PPV was low (27.7%), while NPV was high (99.7%). Heatmap-based localization showed reduced sensitivity (43.8%) despite high specificity (97.5%). These findings demonstrate high specificity and NPV, with lower sensitivity for localization and low PPV in a low-prevalence setting. Prospective multi-institutional studies are required to further validate these diagnostic performance metrics and assess generalizability across diverse clinical settings and imaging protocols.
]]>Authors: Madhini Balasundaram Supraja Perumal
Waste management on university campuses is complicated by the constant change in packaging types, which existing waste-sorting systems cannot recognise unless they are retrained. Open-vocabulary object detectors can identify objects from text descriptions instead of a fixed list of categories, offering a possible solution to this problem. However, it is not known how well this ability survives when such a detector is fine-tuned and deployed on low-power hardware. This paper presents EcoSortBin, a waste-sorting system built on the YOLOE-26 detector and deployed on a Raspberry Pi 4. YOLOE-26 was first fine-tuned on a 1330-image campus waste dataset covering seven classes, with masks generated using the Segment Anything Model, establishing a baseline called WasteYOLOE26-S with 74% top-1 accuracy on known classes; however, this fine-tuning reduces the model’s ability to recognise the same seven classes when they appear in a different dataset or setting. RLPA (RepRTA-Compatible LoRA Prompt Adapters) addresses this by adapting only the text-embedding component of the model using a small set of additional parameters (16,384 parameters, rank 16), leaving the rest of the network unchanged; this restores cross-domain generalisation but reduces top-1 accuracy on known classes to only 15%, which is too low for practical use. To recover this accuracy without losing cross-domain generalisation, frozen-backbone neck fine-tuning (NeckFT) was added, which fine-tunes the feature-combining layers of the network while keeping the main backbone frozen, preserving its pretrained visual–text alignment. Combining RLPA with NeckFT achieved the best balance of the three approaches, with 73.5% top-1 accuracy and a Cross-Domain Generalisation Ratio (CDGR) of 0.2435. To test whether this ability extends to genuinely new categories, the model was further tested on 28 novel categories not seen during training, totalling 840 images. RLPA + NeckFT showed consistent zero-shot generalisation to novel objects with container-like shapes, such as bottles and jars. After quantisation for edge deployment, the model kept its full accuracy ranking and produced a compact 41.8 MB file suitable for the Raspberry Pi 4. These results show that RLPA + NeckFT gives a practical balance of accuracy and generalisation for campus waste detection on low-power hardware.
]]>Authors: Xinran Zhang Yang Sheng Sijie Shen Dongjie Fan Lizhuang Liu
Learning from high-dimensional, low-sample-size (HDLSS) data remains a persistent challenge in machine learning, as models must infer reliable patterns from limited observations while handling an excessive number of variables—a scenario particularly prevalent in biomedical applications. Such data structures render predictive modeling highly vulnerable to erratic optimization and overfitting. To address this challenge, we propose the Global Interaction Network with Dynamic Generalization Control (GINet-DGC), an artificial intelligence (AI) framework that integrates feature-wise structural priors with dynamic generalization monitoring. Rather than directly learning an unconstrained first-layer weight matrix, GINet-DGC generates task-specific weights from multi-view feature descriptors, encompassing latent semantic, global distributional, local topological, and hierarchical representations. This structure-constrained weight generation strategy effectively narrows the feature-interaction search space and acts as an inductive regularizer against noise and redundant molecular features. Furthermore, we introduce an Overfitting-aware Index (OFI) to monitor the training trajectory and effectively identify the generalization saturation point for adaptive termination. Empirical evaluations on eight public real-world biomedical HDLSS gene-expression datasets, using a repeated stratified 5 × 5 cross-validation protocol, demonstrate that GINet-DGC achieves competitive and stable performance against 17 baselines. These findings support the effectiveness of the proposed framework within the evaluated public biomedical HDLSS benchmark setting.
]]>Authors: Camilo Aurelio Velandia Nelson Iván Bedoya Andrés Chiappe David Muñoz-Ballier
This study identifies and characterises the Socratic mediation patterns enacted by MIA, an AI-based conversational agent used in distance higher education. A deductive content analysis was conducted on 737 conversations using six categories: exploration of prior knowledge, contextual adjustment, linkage to experiences, autonomy-oriented prompts, dialogic progression, and verification prompts. The categorical framework achieved full expert content-validity agreement (S-CVI/Ave = 1.00). Contextual Adjustment (77%), Verification Prompts (76%), and Autonomy-Oriented Prompts (74%) were the most frequently observed categories. Sixty of the 64 theoretically possible category combinations occurred in the corpus, and Linkage to Experiences appeared more frequently in personal conversations (33.7%) than in academic conversations (18.3%). The distribution of categories also varied according to conversation length, with longer exchanges containing a broader range of coded dialogic moves. These findings describe the conversational repertoire through which MIA operationalised features associated with Socratic mediation. Because the study did not include independent measures of student satisfaction, learning, engagement, or self-regulation, the results should not be interpreted as evidence of educational effectiveness or causal effects. The study contributes an operational framework for analysing Socratic features in AI–student interactions and identifies directions for outcome-based research.
]]>Authors: Xinru Guo Linze Bai
Accurate mapping of paddy rice is essential for agricultural monitoring, yield estimation, and food security assessment. However, optical imagery is often affected by clouds and spectral confusion, while SAR imagery suffers from speckle noise and weak spatial detail representation. Simple optical and SAR feature concatenation is therefore insufficient for complex agricultural landscapes. To address these limitations, this study proposes ACBE-CroFuseNet, an optical and SAR cross-fusion semantic segmentation network for paddy rice extraction using Sentinel-1 SAR and Sentinel-2 optical imagery in Yancheng, Jiangsu Province. ACBE-CroFuseNet introduces two task-oriented designs for paddy rice mapping. First, an attention cross-fusion module is developed to adaptively model modality contributions and spatial responses between optical spectral–textural features and SAR scattering–structural features. Second, a boundary enhancement module with boundary supervision is introduced to strengthen the delineation of fragmented paddy fields and field edges. Multimodal feature aggregation and multi-scale deep supervision are further used to improve feature utilization and segmentation stability. Compared with UNet++, Swin-Unet, CroFuseNet, and CMFFNet under five-fold cross-validation, ACBE-CroFuseNet achieves the best overall performance. The extracted paddy rice area in Yancheng in 2025 demonstrates the applicability of the proposed method for large-scale crop mapping.
]]>Authors: Sofia I. Hernandez Torres Jennifer Achay Scotty Bolleter James A. Bynum Eric J. Snider
Junctional hemorrhage has a high fatality rate due to how difficult it is to control rapid bleeding from major vessels. The available methods to stop junctional blood loss are prone to placement errors as well as failure during transport and during prolonged field care. On the battlefield, medical imaging with a portable ultrasound can be leveraged for visualization of the underlying tissue and application of compression at the anatomical junction to effectively stop blood flow. In this work, we developed AI models for anatomical landmark tracking using a perfused human cadaver model. These AI models were paired with an end-user clinical application to guide proper placement and compression, improving junctional hemorrhage control on the future battlefield. The trained U-Net semantic segmentation model demonstrated strong performance across predictions for both validation and hold-out, blind subjects. Overall pixel accuracy across the dataset was 98.9% for training and 98.6% for blind subjects. The artery and vein predictions achieved the highest class-specific training intersection-over-union scores, both at 0.73. This segmentation model trained to interpret human tissue provides evidence that ultrasound visualization can help guide compression at anatomical junctions. Future work will focus on improving blind performance for implementation of this AI model into closed-loop control of hardware prototypes, delivering real-time predictions and control.
]]>Authors: Mehdi Chrifi Alaoui Nour-Eddine Joudar Mohamed Ettaouil
Automatic detection of stress in social media text holds promise for supporting digital mental health, but most existing Transformer-based approaches are opaque and computationally demanding. This work presents DR-Transformer, a Dual-Regularized Transformer that combines two complementary mechanisms: (i) a group sparsity penalty (L2,1/L2 elastic net) applied to the query and key projection matrices of every attention head, which encourages whole-row sparsity, producing more concentrated and inspectable attention patterns; (ii) a supervised contrastive loss on the [CLS] projection, which organizes the latent space according to the stress label. The architecture is intentionally lightweight (six layers, eight heads, 256-dim embeddings; ∼9.5 M parameters) and runs entirely on consumer-grade hardware (NVIDIA GTX 1660, 6 GB). Experiments on the publicly available Dreaddit dataset (binary stress classification, 2838 train/715 test segments) compare DR-Transformer against Logistic Regression, BiLSTM, a Standard Transformer of identical architecture, and MentalBERT. Across five seeded runs, DR-Transformer (Full) reaches F1=0.876 (bootstrap 95% CI 0.852–0.898), outperforming the Standard Transformer (F1=0.842; McNemar p<0.001 with Bonferroni correction) and performing comparably to the much larger MentalBERT (F1=0.879; p=0.421). Sparse regularization increases the fraction of near-zero attention weights (below 0.01) from 0.215 to 0.682, while the supervised contrastive loss improves the silhouette score of [CLS] embeddings from 0.312 to 0.483. Dual regularization thus combines accuracy, efficiency, and structurally induced attention concentration in a single model which can be trained without specialized infrastructure. We use the term “interpretable” throughout in this restricted, structural sense—to refer to concentrated and inspectable attention—rather than in the sense of established causal or mechanistic faithfulness; this is only partially and indirectly supported by our token deletion analysis.
]]>Authors: Yuxin Wen Peixiao Fan Zhiyu Mao Fang Chi Chenxuan Zhang Yuhong Lu
Modern power grids require safer and more reliable field operations, yet conventional robots often face limitations in unstructured environments because of rigid pre-programming and weak perception–action coupling. This review examines Embodied Intelligence (EI) as an emerging direction for enhancing power-system field operations. We first evaluate the environmental adaptability of morphological carriers, including quadrupeds, humanoids, and unmanned aerial vehicles, and then define the perception–cognition–execution closed-loop architecture used in this review. Three application domains are then examined. Intelligent inspection focuses on active perception and potential open-vocabulary object detection. Live-line maintenance emphasizes Sim-to-Real methods and shared autonomy, while disaster-response applications involve heterogeneous air–ground robotic coordination. The review also discusses the potential for EI to reduce human exposure to hazardous tasks and influence labor structures, while a regional text-based proxy illustrates differences in policy attention to digital infrastructure. Finally, we analyze major constraints, including hardware endurance under extreme climates, edge-computing latency, foundation-model uncertainty and hallucination, cybersecurity, and safety certification. Overall, EI should not be interpreted as a mature replacement for current utility practice; it is a developing technological direction whose safe deployment will require field validation, standardized evaluation, cybersecurity assurance, and continued human supervisory authority.
]]>Authors: Tomáš Valenta Ondřej Rozinek Josef Horálek
The shift from passive predictive models to autonomous agents capable of tool use and multi-step planning moves the AI safety landscape from prediction error to control failure: small misjudgements become irreversible actions, and risks compound across long horizons and populations of interacting systems. We present a structured review and taxonomy of open scientific problems in agentic AI safety, mapped explicitly onto the EU AI Act and the NIST AI Risk Management Framework. The corpus follows a PRISMA-ScR scoping review, assembled through anchor-based citation chaining and curated reading lists across arXiv, the major machine-learning conferences, and selected security and fairness venues, with a primary March 2026 search cut-off (extended to May 2026 during revision for a small number of high-relevance governance and agentic-safety sources), explicit eligibility criteria, and an analytical distinction between open scientific problems and deployment risks. The taxonomy identifies eight problem families spanning reinforcement-learning policies and language-model planners: goal specification, inner alignment, safe learning and robustness, scalable oversight, interpretability, tool-use security, multi-agent safety, and evaluation and assurance. Mapping these onto the two frameworks shows close alignment for some families and notable absences for others, with multi-agent safety surfacing as a regulatory gap. We add a per-family research roadmap with concrete milestones and a practitioner-facing deployment-posture triage, arguing that progress on inner alignment, interpretability for deceptive-alignment detection, and multi-agent safety would most directly reduce compliance uncertainty.
]]>Authors: Jalal Mehdiyev Ramiz Aliguliyev
Detectors built on transformer language models report near-perfect accuracy on standard fake news benchmarks, which suggests the task is almost solved. We argue that much of this accuracy reflects a confound between writing register and veracity: in common benchmarks, the real class is human-written while the fake class is machine-generated or machine-rewritten, so a detector can separate the classes by recognizing AI writing style rather than by judging truth. To evaluate this, we designed a two-regime evaluation. Phase 1 is the laundering regime, comparing untouched human-real articles against laundered fake articles. Phase 2 is register-controlled: the real class is passed through the same cross-LLM laundering chains as the fake class, so both classes are read in one machine register, and the register cue is no longer available to the detector. We train seven detectors on three datasets and evaluate each frozen detector under both regimes. Under Phase 1, detectors appear robust; under Phase 2, detection on WELFake collapses from about 99% to about 62% AUROC and the largest models approach chance. The effect is benchmark dependent, large on WELFake, mild on IFND and near zero on GossipCop, and it is confirmed by bootstrap testing with false discovery rate control. We recommend register-controlled evaluation as standard reporting practice.
]]>Authors: Daniele Giansanti Andrea Lastrucci
Background: Artificial intelligence-driven digital therapeutics (AI-DTx) are rapidly emerging as a transformative paradigm in healthcare, integrating machine learning, deep learning, and generative AI into digital interventions across diverse clinical domains. Despite rapid growth, the evidence landscape remains fragmented, with heterogeneous methodologies, diverse application contexts, and limited cross-domain synthesis. Aim: This narrative view aims to provide an evidence-informed narrative synthesis of the available secondary literature on AI-driven digital therapeutics, primarily focusing on systematic reviews and meta-analyses, to identify emerging patterns, cross-cutting trends, and future directions across clinical and technological domains. Methods: A narrative synthesis of secondary evidence was conducted, focusing on 23 systematic reviews, meta-analyses, and relevant review articles addressing AI-driven digital therapeutics. The identified literature was analyzed to explore recurring themes across clinical domains, technological approaches, and implementation challenges. Findings were further contextualized through selected recent randomized controlled trials and translational studies to provide insights into emerging clinical applications and real-world perspectives. Results: Across the available literature, AI-driven digital therapeutics demonstrate a broad and rapidly evolving expansion across mental health, chronic disease management, rehabilitation, and behavioral health. The field is characterized by a progressive shift from static, rule-based interventions toward more adaptive systems supported by machine learning, deep learning, and generative AI. A key emerging theme is the role of AI as an enabling layer for personalization, adaptation, and dynamic intervention delivery rather than as a standalone therapeutic modality. Mental health represents the most extensively studied domain, particularly through conversational agents and cognitive behavioral therapy-informed interventions, while other clinical areas are progressively expanding their translational potential. Persistent challenges include methodological heterogeneity, limited long-term validation, and incomplete integration into routine clinical workflows. Discussion: The current evidence suggests a transition toward hybrid human–AI models of care, in which digital systems may support and augment clinical practice through adaptive and data-driven approaches. However, the field remains characterized by fragmented evidence, evolving evaluation approaches, and challenges related to standardization, validation, and real-world implementation. Conclusions: AI-driven digital therapeutics are evolving toward increasingly adaptive and clinically oriented healthcare solutions. Future progress will depend on improving methodological consistency, strengthening long-term evaluation, and supporting responsible integration into clinical pathways to ensure safe, scalable, and meaningful impact.
]]>Authors: Christian Pascal Silva Bouix Vinícius Villa e Vila Marcos Roberto Benso Sergio Nascimento Duarte Carlos Roberto Padovani Roseli Aparecida Francelin Romero Patricia Angélica Alves Marques
The escalating frequency of extreme hydrological events under environmental uncertainty poses a severe socio-economic threat to floodplains such as the Brazilian Pantanal, the world’s largest tropical wetland. Mitigating dynamic flooding and drying cycles is highly challenging due to a critical scarcity of in situ monitoring, leaving flood risks poorly understood. To address these data gaps, this study presents an advanced deep learning forecasting framework that integrates multisource environmental data, fusing satellite-derived precipitation (CHIRPS) and global land data assimilation evapotranspiration (GLDAS) data with historical river gauge telemetry. Multi-layered neural network architectures were optimized and combined with progressive moving average filters (10− and 15−day windows) to capture the complex hydrometeorological patterns of the data-scarce Miranda River Watershed. The optimal deep learning configuration, utilizing a robust two-hidden-layer topology (15 and 60 neurons), consistently outperformed standard baselines. Although purely exogenous data blocks successfully minimized satellite noise and captured seasonal trends (NSE ≥ 0.92), structural underestimation of peak flows was observed. When incorporating the previous day’s streamflow (lag t−1) as a physical anchor, this limitation was noticeably alleviated, increasing both the Nash–Sutcliffe Efficiency (NSE) and Coefficient of Determination (R2) values above 0.99. While this performance surge is driven by the strong temporal persistence inherent to the autoregressive lag, it introduces an operational trade-off by restricting the forecast to a reactive 24 h window. In this regard, an evaluation of the operational forecast horizons revealed that the exogenous deep learning blocks maximize warning lead times, providing a vital tool for proactive civil defense and disaster risk reduction. Ultimately, this multisource framework establishes a methodological foundation for automated decision support systems, providing the high-accuracy streamflow forecasting capability required to support future flood mitigation frameworks.
]]>Authors: Jinyin Bai Wei Zhu Xiangchen Wang Shiluo Guo Zongzhe Nie Tianjin Ni Jinji Zhou Kaiyang Kou Lingxin Xu Yihao Zhong
Post-disaster emergency communication recovery is not merely a link-repair task but a high-level planning problem constrained by service priorities, inter-object dependencies, resource budgets, and time windows. Existing restoration optimization methods generally rely on fully structured inputs, whereas direct large language model (LLM) planning may produce fluent candidates that violate encoded prerequisites, stage-order relations, budget limits, or temporal constraints. To address this challenge, we propose ICG-Restore, an intent-constrained, graph-enhanced LLM planning framework with rule-consistent minimal-edit repair. ICG-Restore transforms mixed restoration requests and structured network observations into task packages that can be checked for validator-level feasibility under an encoded high-level constraint model and evaluated by downstream abstract executors or schedulers. The framework compiles natural-language requests, structured observations, and operational rules into a task-intent object; retrieves task-relevant context from a heterogeneous scenario graph and a restoration knowledge graph; generates stage-wise restoration candidates; and applies bounded local corrections to candidates that violate encoded constraints. In this paper, “minimal-edit” is a descriptive label for a bounded local repair principle that prioritizes less disruptive corrections. Candidates accepted by the validators are evaluated and ranked by a safety-aware agent executor operating in an abstract restoration action space. Experiments on controlled abstract topologies covering three scales, four restoration tasks, and five environmental evolution modes show that ICG-Restore improves validator-level constraint satisfaction and benchmark-estimated recovery utility. Compared with Direct-LLM, it improves CSR and CRS by 1.99% and 24.56%, respectively; benchmark-specific WCTC@5 structural-alignment diagnostic increases by 38.87%.
]]>Authors: Gintautas Dzemyda Modestas Motiejauskas
Images may evoke different emotional responses depending on their content, style, and viewer interpretation. This is particularly important when evaluating works of art, architectural designs, or interior design choices. Visual emotion recognition (VER) models are commonly evaluated using supervised classification metrics such as accuracy, precision, recall, and macro-F1, together with general uncertainty indicators such as entropy, maximum softmax probability (MSP), and top-1/top-2 margin. However, these measures do not indicate whether the model’s strongest competing-emotion predictions remain within the same sentiment group or cross the positive–negative sentiment boundary. This paper proposes the top-2 cross-sentiment rate (CrossRate), a label-free measure for analyzing sentiment-level ambiguity in VER models. CrossRate measures the proportion of samples for which the top-1 and top-2 predicted emotion classes belong to opposite sentiment groups. The measure is evaluated on VER datasets using both standard classification metrics and uncertainty indicators. Experiments on EmoSet-118K show that varying the model’s parameters reduces CrossRate from (22.15±0.45)% to (7.81±0.62)% and increases accuracy from (79.13±0.16)% to (80.10±0.15)%. These changes are not fully reflected by entropy, MSP, or margin, indicating that CrossRate captures a complementary aspect of sentiment-level prediction behavior. The WikiArt case study further demonstrates that CrossRate can be applied when ground-truth emotion labels are unavailable. The proposed measure is applicable to any VER model whose predicted emotion classes can be mapped into positive and negative sentiment groups. The application of CrossRate is illustrated by its use in estimating the emotions of artworks. It offers even non-art experts the opportunity to form an opinion.
]]>Authors: Teng Shi Mingshan Xie Zhenxin Zhao Jifeng Gao
Timely and accurate identification of fish diseases, along with the provision of corresponding treatment plans, is crucial for improving fish welfare and reducing economic losses in aquaculture facilities. To address issues such as low detection efficiency and excessive reliance on manual experience in aquaculture, this paper proposes an innovative diagnostic and treatment method based on the Octopus framework for model optimization. The CBAM (Convolutional Block Attention Module) attention mechanism is introduced into the YOLOv11n backbone to enhance feature extraction, while the Focus-CIoU (Focus-Complete Intersection over Union) loss function is adopted to improve the detection performance for small-target lesions and dense fish schools. Additionally, by integrating wireless sensor networks and the OTDA (Octopus-inspired Therapeutic Decision Agent) bionic agent, an integrated ‘detection-identification-perception-recommendation’ framework is constructed. The results showed that the optimized YOLOv11n model achieved a detection precision of 98.00%, a recall of 94.86%, and an F1 score of 96.40% in detecting ulcer disease, tail rot disease, and red skin disease. Furthermore, the mAP@50 was 97.58%, and the mAP@50–95 was 85.68%. Compared with the traditional YOLOv11n model, the optimized model achieved improvements of 1.39%, 1.48%, 1.43%, 1.4%, and 2.74% in precision, recall, F1 score, mAP@50, and mAP@50–95, respectively. The treatment recommendations produced by means of the Octopus bionic intelligent agent resulted in an average system response time of no more than 53 s, with the maximum observed response time not exceeding 60 s. The identification accuracy of OTDA remained above 87.1%.
]]>Authors: Alfonso Alejo-Ramirez Rogelio Cedeño-Moreno Luis A. Morales Hernandez Juan C. Jauregui-Correa Irving A. Cruz-Albarran
The automotive industry is adopting the Industry 4.0 model to reduce failures in electrical resistance spot welding by using automated non-destructive testing systems. This study presents a real-time machine vision system that has been implemented on an automotive cabin assembly line to detect weld defects in door frames. The system extracts the physical parameters of each spot weld, including nugget diameter and heat-affected zone, to identify failing robots and welding points and prioritise maintenance actions (maintenance-free, predictive, preventive or corrective). Machine learning models were trained using images from 674 cabins collected over 30 continuous hours. YOLOv8 was used for spot weld detection and feature extraction. A neural network with linear discriminant analysis was applied for defect classification, achieving 95.80% accuracy, 95.61% precision, 96.00% recall and 95.80% F1-score. Additionally, a convolutional neural network was developed for maintenance prediction, achieving 94.72% precision, 95.87% accuracy, 93.97% F1-score and 94.42% recall. The results demonstrate the effectiveness of real-time defect detection and reliable maintenance prediction, supporting informed decision-making and efficient resource management.
]]>Authors: Salma O. Ordoñez-Sedano José E. Valdez-Rodríguez Rosa M. Quispe-Siccha
Photoacoustic signal analysis in weakly absorbing media remains challenging because of low signal-to-noise ratios. This work proposes a deep learning framework for classifying particle size and concentration in an indirect absorption configuration. We conducted a comparative study using raw temporal signals, Savitzky–Golay filtering, and time–frequency scalograms via Continuous Wavelet Transform (CWT), and evaluated both 1D and 2D convolutional neural network architectures. Experimental validation was performed using poly(methyl methacrylate) (PMMA) microspheres (6 μm and 15 μm) and hydroxyapatite nanoparticles (<200 nm) at volume fractions as low as 6×10−4%. While raw signals led to unstable training (accuracy ≈ 47%), CWT-based representations significantly improved performance, achieving near-perfect size discrimination and over 96% accuracy in discrete volume-fraction classification. Grad-CAM analysis confirmed that the model identifies physically meaningful regions of the acoustic waveform, ensuring interpretability. The proposed framework was validated under controlled experimental conditions using discrete particle types and predefined volume-fraction classes, providing a foundation for future extensions toward continuous particle characterization. Ultimately, these findings demonstrate that combining time–frequency representations with deep learning provides a robust, physically consistent approach for particle characterization in turbid media, with significant potential for biomedical diagnostics and material analysis.
]]>Authors: Shaoyuan Weng Zongwen Fan Liton Devnath
The liver performs numerous essential metabolic and regulatory functions; however, chronic alcohol consumption remains a major cause of hepatic disease. Consequently, the early and accurate prediction of alcohol-related liver disorders is of considerable clinical importance. A key challenge in this task is the class imbalance present in standard liver disorder datasets, which renders the default classification threshold of 0.5 suboptimal for minority-class prediction performance. To address this issue, we propose a swarm intelligence-enhanced framework that integrates the Butterfly Optimization Algorithm (BOA) with CatBoost. Unlike conventional approaches that either tune hyper-parameters alone or adjust classification thresholds as a post-processing step, the proposed method simultaneously optimizes both CatBoost hyper-parameters and the classification threshold. By incorporating threshold optimization into the BOA objective function, the model adaptively balances sensitivity and specificity without altering the original class distribution. Experiments conducted on two benchmark liver disorder datasets demonstrate that the proposed BOA-driven dual optimization consistently improves performance over baseline classifiers across multiple evaluation metrics. These results demonstrate that joint hyper-parameter and threshold optimization provides an effective and computationally efficient approach for imbalanced liver disorder prediction.
]]>Authors: Benito Samuel López Razo Adrián Trueba Espinosa Farid García Lamont Rosa M. Valdovinos Rosas José Israel Campero Domínguez
The creation and authentication of real users on social media requires the implementation of artificial intelligence-based technologies that can mitigate malicious behavior from automated accounts. This study presents a machine learning-based approach for detecting social bots on Twitter/X, based on the analysis of user profile features and behavioral attributes. Four classification models were evaluated: a neural network (NN), support vector machines (SVM), a random forest classifier (RF), and Extreme Gradient Boosting (XGBoost), using five-fold stratified cross-validation. To improve the performance and robustness of the classification, additional features and data balancing techniques were incorporated. The experimental results show that the neural network achieved the best overall performance, with an average accuracy of 95.6 ± 0.6%, followed by the random forest (95.0 ± 0.6%), the linear SVM (94.1 ± 1.5%) and XGBoost (94.0 ± 1.1%). These results demonstrate that the proposed methodology improves the automated detection of social bots while maintaining the interpretability of the models, which contributes to the development of more reliable and explainable security mechanisms for social media platforms.
]]>Authors: Raiymbek Magazov Kuanysh Abeshev Yernar Shamuratov Fatima Uralova Gulnur Aksholak
Large language model (LLM) backends increasingly generate SSH-honeypot output, yet the literature fixes one backend per system and judges realism by human evaluators, never asking which LLM is fit to play the shell or measuring the catastrophic failure of a backend leaking its own instructions. We fix one hardened unprivileged user scaffold and prompt, vary only the backend across eleven LLMs, and replace the human judge with an objective, prompt-anchored leakage metric. Across a controlled 42-command battery (20 trials each; 8736 responses) and a live adaptive corpus (8626 responses), we score six dimensions: instruction leakage, fidelity, hallucination, latency, verbosity, and stability. Exactly one backend (gemma-4-31b-it) reproduces verbatim leakage in both datasets and is disqualified; the other ten never leak. A deterministic handler layer serves 33 of 42 commands identically across backends, confining model risk to nine generative commands, where flag fabrication ranges 0–95%, latency spans an order of magnitude, and a held-out classifier identifies the backend from one response at 71% versus 9% chance, a fingerprinting risk. We report a per-dimension scorecard rather than a weight-sensitive ranking. Honeypot safety on the privilege boundary is a property of the architecture; realism, speed, cost, and one catastrophic leak are properties of the model.
]]>Authors: Catalin Anghel Andreea Alexandra Anghel Marian Viorel Craciun Adina Cocu Simona Moldovanu Cristian Sandu
Background: LLM-as-a-judge workflows are increasingly used to evaluate open-ended model outputs, but the judge model can itself become a source of error in safety assessment. SafetyJudge-LLM audits local open-weight LLMs as semantic safety judges. Methods: This study reused a fixed set of previously reviewed safety-boundary responses and their hidden reference labels. Two independent human evaluations (R1 and R2) quantified reference-layer ambiguity. Seven local open-weight judge models were evaluated under a common Ollama inference protocol. A paired C6 sensitivity analysis reran llama3.2:3b and qwen3:8b through Hugging Face Transformers. Results: The final judge-output matrix contained 10,612 retained outputs. R1–R2 agreement was 95.45% (Cohen’s κ = 0.612) overall but 47.80% (κ = 0.341) in secondary cases. Several judge models detected more than 90% of confirmed safety-boundary failures, but high detection was not always accompanied by low false-unsafe behavior on control cases. Output-format reliability also varied across models: overall label parseability was 98.11%, while strict JSON schema compliance was 92.55%. The llama3.2:3b schema-failure rate persisted across engines (52.06% under Ollama; 59.60% under Transformers), whereas qwen3:8b maintained complete compliance. Conclusions: SafetyJudge-LLM shows that local open-weight LLMs can support semantic safety judging, but their reliability must be evaluated across multiple dimensions.
]]>Authors: Haruto Takase Masaki Fukunaga Wenwei Yu Jose Gomez-Tames
Background: Transcranial magnetic stimulation (TMS) enables functional brain mapping by linking stimulation (electric field) to motor evoked responses (MEPs), but accurate mapping often requires many empirical samples. Prediction models based only on TMS setup parameters have limited ability to account for subject-specific anatomy, whereas 3D intracranial E-Fields provide richer biophysical information but may suffer from sparse volumetric representations. This study aimed to predict MEPs from individualized cortical E-Field distributions using flattened two-dimensional cortical maps. Methods: TMS–MEP experiments were conducted in seven healthy participants. MRI-derived head models were used to simulate FEM-based intracranial E-Fields for each stimulation condition. Primary motor cortex E-Fields were transformed into two-dimensional cortical maps, and multiple deep learning models were evaluated using nested cross-validation with held-out-subject testing with subject-wise inner validation (HOS-SIV) as a stricter setting and held-out-subject testing with random inner validation (HOS-RIV). Results: ROC-AUC values ranged from 0.710 to 0.811 and from 0.840 to 0.872 across muscles under the held-out-subject testing with subject-wise inner validation (HOS-SIV) and held-out-subject testing with random inner validation (HOS-RIV) settings, respectively. Under the primary HOS-SIV setting, conventional classifiers and CNN-based models showed broadly comparable performance, whereas CNN- and ResNet-based models tended to perform better under the HOS-RIV setting. Predicted spatial MEP maps partially reproduced the measured response distributions. Conclusions: The presence of TMS-induced MEPs can be predicted from flattened maps of subject-specific cortical E-fields, providing proof-of-concept support for future approaches to improve the efficiency of TMS motor mapping.
]]>Authors: Sumin Park Mu-Gyeong Gong Sang-Jae Park Sangseok Yun Il-Min Kim Jeehyun Kim Jae-Mo Kang
Object removal is a widely used AI-eraser operation in photo editing and privacy protection, yet conventional workflows make the user paint the removal region by hand—a burden that is most acute when an image contains several objects of the same class and only one is to be erased. We present two pipelines for maskless selective removal that delete a designated object among many, using only a target ID, box, or point. Both are built around a selection mechanism: the detected candidates are indexed, and only the mask of the designated target—combined by a logical-OR when several are chosen—is constructed and inpainted, unlike the standard use of an inpainting model, which merely fills a mask that is already given. On the multi-object GQA-Inpaint benchmark, we compare the generative YOLO–SAM–SDXL (YSS) and the lightweight YOLO/MobileSAM–LaMa (YML) against PowerPaint and Inpaint-Anything. With no mask provided, the proposed pipelines select and remove the target far more accurately and reliably than the baseline (for YML, Selection-IoU 0.626 vs. 0.327 and residual object score 0.228 vs. 0.646), while their full-image quality stays within the baseline range. YML is the most reliable remover and far faster (1.31 s per image on GPU vs. 29.60 s for YSS, and 32× faster than a diffusion pipeline on CPU), whereas YSS reaches the most natural perceptual quality at the cost of a higher residual. The two are thus complementary options, chosen according to the desired speed and restoration character.
]]>Authors: Muhammad Talha Sajid Ahmad Kamran Malik Nafees Qamar Hasnain Abdullah Aleem Ahmed
In biomedical research and clinical practices, Graph Neural Networks (GNNs) are playing an increasingly important role and have been applied to the problems of disease pathway detection, gene–disease relation prediction, etc. They show great potential for biomedical predictions; however, there are interpretability issues when used on complex datasets like gene expression data. Current explainability methods such as GNNExplainer are designed to explain individual instances, not the whole network. The absence of transparency hinders trust and limits the clinical/biomedical implementation of GNNs. Additionally, more interpretable models like GNN-SubNet and XGDAG do not fulfill the expectation of a clear picture for the entire network. This research addresses the limitation of the network-wide explainability of GNNs by introducing a GNN-based BioGraphEX model that incorporates interpretable methods at two levels, instance-level and network-wide level, such as gradient-based methods and SHAP (Shapley Additive Explanations). Using the GSE25097 biomedical dataset, the model achieves an accuracy of 85% and an F1 Score of 0.82, surpassing baseline methods in both predictive performance and interpretability. These results address the limitations of existing models like GNN-SubNet and XGDAG by providing both instance-level and network-wide insights. Metrics like Explanation Fidelity (83%) further validated the robustness of the explanations.
]]>Authors: Yue Ling Yucheng Cao Jianghong Yu
High-temperature creep data for aero-engine fasteners are essential for life assessment and structural safety, but such records are often archived as heterogeneous tables, reports and figures with weak cross-record associations. This study develops a traceable knowledge-graph retrieval-augmented generation (KG-RAG) workflow for internal GH2132 fastener creep data. Seventeen GH2132 creep specimens with different geometries were organized into a standardized document archive, parsed into source-linked table and text fragments, and converted into structured entities and relationships. Iterative prompt templates were used to check extraction completeness, repair missing fields and preserve provenance without fine-tuning the underlying language model. The resulting Neo4j knowledge graph links material, specimen geometry, service condition, creep measurement and source evidence, while LightRAG retrieves vector chunks, entity contexts and relationship contexts through round-robin merging. The system was evaluated using 20 seed-engineering questions across five retrieval modes (Naive RAG, Local, Global, Hybrid and Mix), yielding 100 query-evaluation records. The questions covered simple, comparative, multi-constraint, out-of-knowledge-base and ambiguous queries. Mix achieved the strongest RAGAS profile, with answer correctness, answer similarity, answer relevancy, context precision, context recall and faithfulness scores of 0.85, 0.83, 0.79, 0.88, 0.87 and 0.85, respectively. Hybrid may be more practical for routine queries, although this qualitative interpretation was not supported by systematic latency benchmarking. A separate assessment by 15 domain experts gave a weighted score of 94/100. These results show that a source-linked KG-RAG workflow can improve the reuse, grounding and auditability of heterogeneous creep records within a bounded engineering dataset.
]]>Authors: Pablo Rivas Liang Zhao
Artificial intelligence (AI) marketing systems increasingly plan campaigns, generate messages, assess performance, and revise outputs with limited human input. In such multi-agent systems (MAS), commercial optimization pressure may produce ethical compliance drift: gradual movement from acceptable persuasion toward detector-defined manipulation-risk signals. We examine this problem in a controlled OpenClaw simulation of a three-role AI marketing team. The study varies optimization pressure across baseline, moderate-pressure, and high-pressure operating regimes while holding the model backend, prompt pool, agent roles, and monitoring process constant. Optimization pressure significantly affected detector-defined manipulation-risk scores, and the observed pattern was non-monotonic. The moderate-pressure condition produced the highest observed mean, but moderate and high pressure were not statistically distinguishable in the pairwise comparison. These results reflect detector-based risk indicators in simulated marketing-agent outputs, not direct evidence of consumer harm, deception, or real-world behavioral manipulation. The findings support pressure-aware auditing as a standards-informed monitoring practice consistent with IEEE value-sensitive design principles, while also identifying the need for human annotation, hybrid detectors, and real-user validation before stronger governance claims are made.
]]>Authors: Yuv Raj Pant Haitham Y. Adarbah Afzel Noore Dunren Che Aden Ahmed Robert Ayala
Large Language Models (LLMs) show promise for supporting formative assessment by generating feedback on students’ written mathematical reasoning. However, practical use in educational settings remains constrained by the need to manually curate representative few-shot exemplars for prompt construction. This study examines whether unsupervised semantic clustering can automate few-shot exemplar selection for LLM-generated formative feedback in middle-school mathematics. As a controlled methodological feasibility study, we generated and refined 100 exam-realistic constructed responses for a Grade 7 inequality task aligned with middle-school mathematics standards. Student responses were embedded using Sentence-BERT, projected into a lower-dimensional space using Uniform Manifold Approximation and Projection (UMAP), and clustered with Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) to identify dominant reasoning patterns and ambiguous responses. Representative centroid and boundary exemplars from the resulting clusters were then used to construct few-shot prompts for the Llama 3.3 70B model, which generated feedback for the remaining 93 responses. Six independent mathematics instructors evaluated the AI-generated feedback using a structured 0–5 usability rubric. Across all instructor evaluations, 538 of the 558 instructor ratings (96.42%) were 3–5, representing feedback ranging from fair, requiring moderate edits, to excellent, ready to send. More specifically, 482 of the 558 instructor ratings (86.38%) were scores of 4–5, indicating feedback requiring no edits or only minor revisions. The remaining 20 ratings (3.58%) were scores of 0–2, while 13 unique feedback messages received at least one low rating. Across all instructor evaluations, 20 of 558 ratings (3.58%) were assigned scores of 0–2, while 13 unique feedback messages received at least one low rating. Qualitative analysis of low-scoring cases revealed recurring failure modes, including hallucinated completeness in concise solutions, failed arithmetic verification, and false logic flagging for atypical reasoning patterns. These findings suggest that clustering-based exemplar selection may reduce manual prompt-engineering effort while supporting usable LLM-generated formative feedback in a controlled mathematics setting. However, the present study does not compare clustering against alternative exemplar-selection strategies, and therefore conclusions should be interpreted as evidence of feasibility rather than comparative superiority.
]]>Authors: Xiaoli Hu Yang Shen
Social and behavioral research often needs to examine policy shocks, information interventions, platform-mediated attention, and governance feedback, but direct experiments on real populations are constrained by ethical risks, intervention costs, and limited repeatability. This study proposes Multi-Agent Social Simulation (MASS), a protocolized form of large language model-driven agent-based modeling (LLM-driven ABM) designed as a low-risk, repeatable, and auditable pre-experimental simulation method for quantitative research. MASS embeds LLMs in an agent-based modeling (ABM) framework and uses role settings, round-based scheduling, information control, background-rule control, structured outputs, harness checks, reason-action logs, and replication manifests to transform open-ended language generation into recordable, checkable, and statistically analyzable agent-round observations. The method is evaluated through the New Jersey–Pennsylvania minimum wage natural experiment, the 2016 UK Brexit digital campaigning context, and the 2023 Zibo barbecue tourism public-opinion event. Results show that protocolized LLM-driven ABM can generate analyzable and empirically assessable outputs across policy-shock, information-intervention, and governance-feedback scenarios. The strongest evidence concerns rule-shock identification, declining undecided share under targeting, and mechanism-chain consistency among governance response, public sentiment, and behavioral intention. MASS is not a substitute for real-world experiments or causal inference; it is a pre-experimental simulation method for mechanism rehearsal, risk identification, counterfactual comparison, and research design preparation.
]]>Authors: Omneya Attallah
Potato leaf disease poses a significant threat to global food security, causing substantial crop losses that jeopardise agricultural productivity and farmers’ livelihoods worldwide. Existing automated detection frameworks suffer from several persistent limitations, including over-reliance on controlled benchmark datasets, narrow disease class coverage, exclusive use of spatial feature representations, absence of feature selection, and dependence on single-architecture end-to-end pipelines. To address these limitations, this paper proposes ComPo-Net, a novel lightweight ensemble framework that integrates three efficient CNN architectures—ResNet18, ShuffleNet, and MobileNetV2—for nine-class potato leaf disease detection and classification. Deep features are extracted from three intermediate layers of each network, with the Discrete Wavelet Transform applied for dimensionality reduction and cross-network fusion of the higher-dimensional layer features, capturing spectral–spatial information that purely spatial approaches cannot provide, while the remaining layer features are directly concatenated across networks. One-way Analysis of Variance (ANOVA) feature selection is subsequently applied to retain the most statistically significant features from the combined multi-scale, multi-network representation, and seven machine learning classifiers are systematically evaluated to identify the optimal classification strategy. The framework is assessed on a merged dataset of three publicly available benchmarks spanning both controlled and uncontrolled imaging environments, constituting a nine-class evaluation setting not previously addressed at this scale in the literature. ComPo-Net achieves an accuracy of 96.32%, an F1-score of 93.87%, an MCC of 0.9350, and AUC values exceeding 0.993 across all nine classes with Cubic SVM as the best-performing classifier. When compared against methods evaluated on the seven-class uncontrolled-environment dataset—the closest available task setting to ComPo-Net’s nine-class merged benchmark—ComPo-Net surpasses the best-performing comparable method by a margin of 6.45 percentage points, demonstrating the effectiveness of multi-scale ensemble feature extraction combined with spectral–spatial representation and principled feature selection for robust potato leaf disease detection under diverse real-world conditions.
]]>Authors: Ahmed El Badaoui Abdellah Ezzati Said Ben Alla Manal Hilali Hicham Ben Alla
Self-supervised graph collaborative filtering suffers from two geometric pathologies. In Dimensional Collapse, embeddings quietly collapse into a low-rank subspace, a well-documented but poorly solved problem. Semantic Collapse is more complex: stiff L2-squared orthogonalization penalties that are supposed to push the embeddings apart end up ripping through the heavy-tailed community overlaps that contain the collaborative signal. Earlier studies attempted to mitigate sparsity by injecting static noise or structural perturbations, but such interventions did not pinpoint the root cause, i.e., the distortions in the global covariance geometry itself. In this paper, we propose a Huber-Contrastive Graph Convolutional Network (HCGCN) that combines a spatial message-passing backbone with an O(1) contrastive augmentation overhead and a Relation-Aware Dual-View Gated Contrastive Network. The main novelty is a Huber Covariance Whitening module that imposes a geometry-aware threshold on the cross-correlation matrix of augmented views—below the threshold, the gradients follow an L2 penalty (enforcing uniformity); above it, the penalty flattens to L1 (protecting genuine semantic clusters from gradient explosion). This theoretically motivated dual-regime penalty actively preserves the macro-semantic topology of the graph while aggressively stamping out spurious noise correlations. The HCGCN is evaluated on the Yelp2018, Amazon-Book, and MovieLens datasets and performs significantly better than state-of-the-art baselines like LightGCN, SGL, SimGCL, and NESCL, especially under severe cold-start settings where covariance regulation proves most critical.
]]>Authors: Seoljoo Kang Wanju Huang
While computational thinking (CT) is increasingly vital in K-12 education, teaching it through text-based coding remains challenging for teachers. To address this gap, this study presents and evaluates a self-paced professional development (PD) module, “Let’s Code with GenAI,” created for K-12 teachers to enhance text-based coding and CT. Using a one-group pretest-posttest design, 34 pre-/in-service teachers completed the module in 2025, engaging with instructional videos and hands-on coding activities using Micro:bit and MakeCode (v11.3.22), with a GenAI assistant providing explanations and debugging support. Pre- and post-intervention data were collected using the Teacher Beliefs about Coding and Computational Thinking (TBaCCT) scale and a coding/CT assessment. Posttest scores were higher than pretest scores on teaching efficacy and value beliefs (p < 0.001 for both), as well as in coding self-efficacy (p < 0.001) and CT self-efficacy (p = 0.015). Coding/CT assessment scores were also higher at posttest (p = 0.040). No statistically significant correlations were found between self-efficacy and performance measures. Overall, the findings offer preliminary insights into participants’ post-intervention outcomes in a GenAI-supported, self-paced PD module, including self-efficacy, value beliefs, and coding/CT performance, while underscoring the need for future controlled studies.
]]>Authors: Rahul Razdan Dmitri Mironov Janika Leoste Mohsen Malayjerdi Mauro Bellone Raivo Sell
Digital Artificial Intelligence (AI), exemplified by Large Language Models (LLMs) such as ChatGPT, has achieved remarkable progress across a wide range of applications, driven not only by advances in algorithms but also by the emergence of a shared research ecosystem built upon commodity computing platforms, standardized software frameworks, open-source models, benchmark datasets, cloud infrastructure, and broadly accessible educational resources. In contrast, Autonomous Vehicles (AV), AI systems that perceive, reason, and act in the physical world, have advanced more slowly despite substantial public and private investment. Progress remains constrained by fragmented research and educational infrastructure that limits reproducibility, interoperability, scalable validation, and workforce development. This paper surveys the current state of the AV research ecosystem, including hardware platforms, autonomy software stacks, datasets, simulation environments, digital twins, testing and validation frameworks, and educational programs. Drawing lessons from the evolution of Digital AI, the paper identifies key gaps in accessibility, standardization, integration, and openness across the AV technology stack and outlines opportunities to develop shared research testbeds, modular open platforms, interoperable software and data ecosystems, common benchmarks, and interdisciplinary educational programs that can accelerate autonomous vehicle innovation. Finally, the paper provides a framework for evaluating AV research and educational infrastructure which identifies priorities for future investment.
]]>Authors: Armin Stein Umut Volkan Kizgin Niklas Waldmann Bjarne Käberich Souhaiel Ben-Salem Aaron Dlugosch Ajay Kumar Thakur Thomas Vietor
Large Language Models (LLMs) show growing potential for transforming natural-language engineering information into formal model-based artefacts. However, their reliability for SysML-v2 artefact generation remains insufficiently understood, particularly regarding prompt-strategy selection, model-dependent variability, and task-specific performance. This paper proposes a prompt-strategy-driven method for LLM-assisted SysML-v2 artefact generation in Model-Based Systems Engineering. The method is grounded in a systematic literature analysis and a Harvey-Balls-based assessment of existing approaches, which reveal gaps in reproducibility, prompt evaluation, and formal modelling support. The proposed workflow integrates input preparation, prompt-strategy selection, LLM selection, artefact generation, syntax validation, and quality evaluation. It is evaluated using a traction battery system case study across three representative modelling tasks: requirements generation, block definition modelling, and state-machine modelling. Four LLMs are compared using five prompt strategies and assessed through F1-score analysis and LLM-as-a-Judge evaluation. Within the investigated traction-battery case study, the observed artefact scores differed across prompt strategies and modelling tasks, and no single strategy achieved the highest observed score across all tasks. Overall, the findings indicate that LLMs can support early-stage SysML-v2 modelling as human-in-the-loop assistants, while expert validation remains necessary to ensure syntactic correctness, semantic consistency, and domain validity.
]]>Authors: Lavdim Menxhiqi Galia Marinova
In a closed-inventory large language model (LLM) system such as Online-CADCOM, which recommends engineering tools from a verified inventory, we measure inventory non-compliance, that is, a mention-level event in which the model recommends a tool not present in the verified inventory. We use this inventory-relative sense of hallucination throughout: an out-of-inventory mention may be a fabricated tool or a real commercial tool absent from the curated inventory, so the metric reports inventory non-compliance rather than factual fabrication. We evaluate a three-mechanism mitigation stack consisting of database-grounded context injection, fixed vocabulary constraints, and enforced JavaScript Object Notation (JSON) output across three commercial LLM providers (OpenAI, Anthropic, Google), two model generations, and two output modes (standard and reasoning), totaling 6912 Application Programming Interface (API) calls over 12 configurations. Under a recall-equalized detector adopted as the primary metric, the inventory non-compliance rate, which we denote the hallucination rate (HR) following common usage, decreases from roughly 69–80% to 4–13% under the full architecture. The cross-provider average is similar across the two generations tested (8.5% Generation 1 (Gen1), 6.9% Generation 2 (Gen2)), although per-provider directions diverge. We also examine the C3 configuration, in which only JSON output enforcement is active without grounding. A naive detector reports a large hallucination increase over the unconstrained baseline (+10.1 percentage points (pp) Gen1, +15.1 pp Gen2), but we show this gap is largely a detection-format artifact: structured JSON fields make out-of-inventory tools easy to extract, whereas the same real tools are frequently missed in free text. Under a recall-equalized detector the gap narrows to +2.6 pp (Gen1) and +4.8 pp (Gen2) and remains statistically significant only for two current-generation models, indicating a small, current-generation effect rather than a universal one. Reasoning-mode models provide no statistically significant improvement under architectural constraints. A frequency-weighted audit shows that the majority of remaining out-of-inventory mentions correspond to real engineering tools absent from the platform’s inventory. Under the full architecture, roughly half of responses (pooled Pany≈49.5%) still contain at least one such mention, indicating that handling unseen tools remains an open challenge for closed-inventory recommendation systems. Our evidence comes from a single engineering platform with four related electronic-design and power-electronics domains, so the findings characterize this setting rather than recommendation domains in general.
]]>Authors: Andreea-Ramona Treteanu Horațiu Herdeș Gheorghe Cobuz Matei Niță Ioana Vișoiu Alexandra Hoștiuc Matteo Gregorini Lorenzo Lorusso Carmen Adella Sîrbu Ana Maria Alexandra Stănescu
Access to neurological care remains profoundly unequal worldwide, driven by the increasing burden of chronic and neurodegenerative disorders and a persistent shortage of specialist neurologists. Tele-neurology has emerged as a promising strategy to improve access to neurological expertise, while recent advances in artificial intelligence (AI) have expanded its capabilities beyond remote consultation toward data-driven diagnosis, monitoring, and clinical decision support. This narrative review critically synthesizes current evidence on the evolution of tele-neurology and the emerging role of AI across multiple domains of neurological care. The review examines AI-enhanced diagnostic applications, including neuroimaging, electroencephalography, and digital biomarkers, as well as AI-assisted remote monitoring, patient engagement, and predictive analytics in disorders such as stroke, epilepsy, Parkinson’s disease, multiple sclerosis, dementia, headache disorders, and neuromuscular diseases. Available evidence suggests that tele-neurology can achieve clinical outcomes comparable to in-person care in selected settings while improving accessibility, continuity of care, and patient satisfaction. AI applications have shown promising early results in image interpretation, automated EEG analysis, remote disease monitoring, and individualized risk prediction. Despite these advances, important challenges remain. Limitations include constraints in remote neurological examination, evidence gaps regarding clinical validation, unequal access to digital infrastructure, data governance and cybersecurity concerns, fragmented regulatory frameworks, and difficulties integrating AI into routine clinical workflows. Furthermore, the clinical maturity of AI applications varies substantially, with many systems remaining at the proof-of-concept or early validation stage. To make this variation explicit, we introduce a transparent, criteria-based three-tier classification of the clinical practicability of AI applications, graded by strength of evidence, degree of validation, and real-world integration. Tele-neurology is increasingly evolving toward hybrid models that combine in-person neurological assessment with digitally enabled longitudinal monitoring and AI-supported decision tools. Future progress will depend on rigorous clinical validation, equitable implementation strategies, integration into healthcare systems, and the development of multimodal AI approaches that combine clinical, imaging, electrophysiological, and digital biomarker data to support personalized neurological care.
]]>Authors: Ziqing Xia Xiazijian Zou Chengguang Liu Ronghua Shi Chao Hu Xiaohang Zhou Ying Fu
Logo detection is a critical technology for safeguarding corporate intellectual property and maintaining cyberspace security by identifying trademark infringements and mitigating phishing risks. While contemporary logo detection algorithms have demonstrated strong performance in natural scene images, they frequently struggle with web-based content due to the structural dependencies between logo icons and surrounding textual elements. Such complexities often lead to sub-optimal detection performance. To bridge this gap, we present HyperLogoDet, a framework specifically engineered for web logo detection that integrates a text detection module with a dynamic hypergraph filtering mechanism to reduce interfering textual noise. Specifically, our approach first employs a specialized text detector to identify potential textual regions. Subsequently, dynamic hypergraph filtering is used to distinguish distracting text from text or icon components that should be preserved. Final logo identification is then performed on the refined, masked image. To evaluate the proposed framework, we introduce three WebLogo benchmarks: WebLogo-500, WebLogo-1000, and WebLogo-1500. Empirical evaluations on these benchmarks show that HyperLogoDet improves logo detection accuracy under textual interference while maintaining competitive inference efficiency. We also analyze threshold sensitivity and discuss practical limitations for future improvement.
]]>Authors: Rüdiger Heimgärtner
Automotive SPICE® (ASPICE) conformance activities require systematic discovery and evaluation of project documentation against standardised process indicators, consuming an estimated 40–60% of assessment preparation time with high extraneous cognitive load. This paper presents Smart Evidence Management (SEM), a principled Human–AI Collaboration (HAIC) approach grounded in Cognitive Load Theory, Trust in Automation, and Situation Awareness theory. SEM defines four design principles—AI-assisted discovery, expert-validated judgement, transparent reasoning, and full data sovereignty—as a novel generalisable HAIC pattern for documentary evidence evaluation in regulated professional domains. By reducing evidence-hunting effort, SEM operationalises the Plan-Do-Check-Act (PDCA) continuous improvement principle, enabling iterative conformance gap detection throughout development rather than only at formal assessment events. The Smart Evidence Manager, a prototype instantiation for ASPICE Process Assessment Model (PAM) v4.0, implements a two-stage hybrid symbolic–neural Retrieval-Augmented Generation (RAG) architecture with fully local Large Language Model (LLM) inference. An expert agreement study (n = 120 findings, five intacs® certified assessors) yielded a combined acceptance rate of 90.0% (95% Confidence Interval (CI) [83.3%, 94.2%]). A beta-test study is under recruitment (target n ≥ 30 practitioners) to quantify cognitive augmentation effects, usability, and trust calibration. Cultural dimensions of trust calibration are discussed, extending SEM to intercultural deployment contexts.
]]>Authors: Ibrahim M. H. Rahman Osama Rehman Aisha Ajmal Simon Jigwan Park
Top-down visual attention models are essential for task-driven object detection; however, many existing approaches do not effectively integrate multiple sources of high-level guidance such as scene context and target-specific information. This paper proposes a computational framework that combines contextual information and target object features to dynamically modulate low-level visual features for improved attentional selection. The proposed model consists of three key components: (i) a contextual weighting module that learns feature weights optimised using Particle Swarm Optimisation and predicted through a hetero-associative neural network; (ii) a target-aware attention module that estimates feature importance based on target-specific characteristics derived from low-level feature distributions; and (iii) a recognition module that performs region classification using a Naïve Bayes classifier. Experiments conducted on seven challenging datasets with diverse objects and cluttered backgrounds demonstrate that integrating contextual and target-specific information improves detection performance compared to using either source independently. The primary objective of this work is to investigate the complementary and synergistic effects of combining contextual information and target object knowledge within a bottom-up saliency framework. The goal is not to propose or claim a state-of-the-art top-down visual attention model, but rather to demonstrate that integrating multiple sources of guidance can enhance saliency-based target detection under varied visual conditions.
]]>Authors: Daniel Cristóbal Andrade-Girón William Joel Marin-Rodriguez Marcelo Gumercindo Zuñiga-Rojas Abrahan Cesar Neri-Ayala Edgar Tito Susanibar-Ramírez Miguel Angel Aguilar-Luna-Victoria
Student stress is a significant mental health issue in educational settings; therefore, developing reliable, calibrated, and interpretable predictive models can support the classification of observed stress levels. This study analyzed the public Student Stress Factors dataset, comprising 1100 records, 20 predictors, and one target variable, using a supervised machine learning pipeline designed to reduce information leakage. The pipeline included stratified data partitioning, encapsulated preprocessing, nested cross-validation restricted to the training data, and independent holdout evaluation. Nine ensemble and boosting algorithms for tabular data were compared: AdaBoost, Gradient Boosting, Random Forest, Extra Trees, Bagging, Voting, Stacking, XGBoost, and LightGBM. Model performance was assessed using key discrimination and calibration metrics, together with the nonparametric Friedman test for statistical comparison. Gradient Boosting achieved the best average performance in nested cross-validation, with an accuracy of 89.55 ± 3.16%, F1-weighted of 89.54 ± 3.17%, MCC of 0.845 ± 0.047, and ROC-AUC weighted of 98.59 ± 0.92%. XGBoost and LightGBM showed comparable performance. In the independent holdout set, the final calibrated model maintained robust predictive performance, achieving an accuracy of 0.8818, F1-weighted of 0.8818, MCC of 0.8237, and ROC-AUC weighted of 0.9861. Although the overall results indicate stable and high predictive performance, the Friedman test did not identify statistically significant differences among the algorithms, χ2 = 10.953, p = 0.204. Therefore, model selection should consider not only predictive accuracy but also computational efficiency, calibration, interpretability, and implementation feasibility. Despite the internal stability of the pipeline and satisfactory holdout performance, the public and cross-sectional nature of the dataset limits causal inference and model transferability. Consequently, external and prospective validation is required before integration into institutional early warning systems.
]]>Authors: Sanjay Basu Parth Sheth Bhairavi Muralidharan John Morgan Rajaie Batniji
Clinical artificial-intelligence systems are starting to act across a course of care, not to answer one question at a time. Their safety is checked by methods that sample the input space: a test suite tries some inputs, a language-model reviewer reads some cases, a physician panel audits some cases. A sampling check can pass a safety rule and still miss the rare input that breaks it, such as a documented obligation dropped several encounters later. This study measures that gap and releases CIV-Bench, a public benchmark of 832 clinical rule sets with safety properties across eight whole-person domains, in single-encounter and longitudinal forms, plus a computational stress tier, each with independently established ground truth. We compare formal verification, which uses a satisfiability-modulo-theories (SMT) solver to check every possible input at once, against the methods used in practice: random unit testing, language-model judges, and a blinded physician panel. Formal verification detected all 612 violations, raised no false alarm, and returned no unsound verdict; for each item it returned either a proof that the rule holds over every input or one concrete input that breaks it. A frontier language-model judge matched this detection, but it returned a pass rate over sampled cases rather than a guarantee, at three orders of magnitude more compute per item. The general open-weights judge returned unsound verdicts on the computational stress tier; the medically fine-tuned judge was unsound far more widely, collapsing on the longitudinal properties despite strong single-encounter medical detection, so medical fine-tuning did not close the gap. Unit testing and the physician panel missed the deep, cross-encounter violations that hold a course of care together. Formal verification is set apart not by a higher detection rate but by the kind of evidence it returns: a proof over the whole input space, a replayable counterexample, or an explicit statement that it cannot decide. The guarantee holds for the decisions placed in this layer, and it depends on the safety rule being specified correctly.
]]>Authors: Jelena Čulić Gambiroža Ana Čulić Kristina Medić Ana Grubišić
This study investigates the application of Artificial Neural Network (ANN), Long Short-Term Memory network (LSTM) as a representative of Recurrent Neural Network (RNN), and one-dimensional Convolutional Neural Network (1D-CNN) for time series prediction, demonstrated through a use case of low-cost gas sensor readings from transient signals. Despite the widespread use of these architectures in IoT forecasting applications, there is a lack of systematic comparative studies that evaluate their performance under identical experimental conditions, particularly in energy-constrained sensing scenarios. The primary objective is to evaluate the trade-offs between model accuracy, computational cost, and memory requirements under energy-efficient data acquisition scenarios. A comprehensive experimental analysis was conducted using 186 recorded transient samples, where all models were trained and evaluated under consistent preprocessing, identical data splits, and uniform hyperparameter settings. Performance was assessed using RMSE, MAE, R2, training time, and model size as key evaluation metrics under varying input sequence lengths. The results show that the LSTM model achieved the highest accuracy, with an RMSE of 3.69%, R2 of 0.85 and scaled MAE of 0.04, effectively capturing long-term temporal dependencies. The 1D-CNN exhibited a balanced compromise between accuracy and training efficiency, while the ANN provided the shortest training time but lower overall performance. Reducing the number of input readings from 186 to as few as 10–20 resulted in only a 2–4% increase in RMSE, with model size reductions of up to 50%, making such configurations particularly suitable for edge or embedded IoT devices. The findings demonstrate that artificial neural networks can maintain high prediction accuracy even under reduced data conditions, contributing to the development of low-power, resource-efficient sensing systems for intelligent and distributed IoT environments.
]]>Authors: Xi Xiao Youchen Fan Ming Lei Liu Yi Qichen Wang Shengliang Fang
Large-scale spectrum monitoring infrastructures generate high-dimensional spectral time series, providing a critical data foundation for proactive spectrum management, anomaly detection, radio environment awareness, and interference-aware decision-making. In complex electromagnetic environments, real-world deployments are highly nonstationary and frequently affected by unexpected interference, which substantially degrades the predictability of spectrum dynamics and the reliability of downstream spectrum sensing and management systems. Consequently, classical linear forecasting methods and generic deep sequence models often generalize poorly from clean training conditions to interference-corrupted scenarios, as jamming patterns distort the latent representations used for future-spectrum forecasting. This study focuses on multivariate spectrum forecasting, where the objective is to predict multi-step future amplitude or power distributions across all frequency bins from a historical observation window. To address this limitation, we propose DS-SpecIT, a Decomposed Spectral Inverted Transformer for interference-aware spectrum forecasting. Unlike generic long-term forecasting models that mainly minimize average prediction errors, DS-SpecIT is specifically designed to handle structured electromagnetic interference. Its novelty lies in the integration of spectral tokenization, inverted attention over frequency tokens, an interference-aware dual-scale objective, and orthogonality-based latent feature separation. These components enable the model to jointly preserve global spectral trends and reduce local errors inside interference-affected time–frequency regions. Using publicly available spectrum measurements, we establish evaluation protocols under both clean and synthetic-jamming settings. Experiments show that DS-SpecIT maintains competitive clean setting accuracy while achieving stronger global and local robustness under structured interference.
]]>Authors: Zhaobin Li Mingliang Mo Chenchong Du Zhanzhen Wei
Federated learning lets hospitals train shared diagnostic models without exchanging patient images, yet the updates they exchange each round can be inverted to reconstruct training images. Homomorphic encryption (HE) protects these updates, but encrypting an entire model with CKKS inflates communication and computation to impractical levels for cross-silo medical use. We present PASHE-FL, which exploits the structure of personalized federated learning: the client-specific classifier head stays local and is never uploaded, so encryption need only cover the shared backbone. The server ranks backbone coordinates by importance from the public global model and selects the same top-ρ set for all clients, avoiding mask negotiation; these are encrypted with CKKS, the remaining coordinates are quantized, and the encrypted fraction is annealed over training. On four medical image-classification tasks, PASHE-FL matches the personalized FedPer baseline within about one accuracy point while cutting per-round uplink by roughly 7.1–7.6× and encryption time by about 8× relative to full-model HE; this accuracy comes from personalization, not encryption. Under a gradient-inversion attack, encrypting only the top 5–10% most important coordinates collapses reconstruction quality, whereas encrypting random coordinates does not. PASHE-FL offers an empirical privacy–cost trade-off under the stated threat model rather than a formal privacy guarantee.
]]>Authors: Gulnaz Zakariya Aiman Moldagulova Nor’ashikin Ali
Payroll-anchored retail borrowers—individuals whose monthly remuneration is routed into an account at the lending institution through a salary-project arrangement—constitute the volume backbone of unsecured consumer lending in Kazakhstan, generating the largest origination flow, the lowest realized default rate, and the majority of the systemic regulatory and capital sensitivities of second-tier banks. Payroll anchoring also changes the lender’s information set, which motivates a study of how that advantage translates into model performance and borrower outcomes. We design and internally validate an explainable hybrid artificial-intelligence framework stratified by client tenure into two production models: a Weight-of-Evidence (WOE) logistic-regression scorecard for new salary-project applicants, and a hybrid scorecard for repeat applicants, in which a stacked ensemble of LightGBM, CatBoost and a multi-head self-attention neural network contributes a single WOE-encoded predictor to a second-stage L2-regularized logistic regression. The hybrid recovers a substantial share of the ensemble’s discriminatory lift while preserving an auditable, monotone scorecard at the point of decision, and isotonic recalibration restores the predicted probabilities of default to the empirical bad-rate scale required for IFRS 9 expected-credit-loss accrual and risk-based pricing. We report discrimination, calibration and stability evidence under a strict anti-leakage protocol and set out the structural preconditions under which the architecture transfers to other emerging-market payroll-anchored portfolios. We are explicit about scope: a true out-of-time validation and a full group-conditional fairness audit are identified as required next steps rather than claimed here. The contribution is a reproducible, interpretable scoring design that exploits payroll visibility while retaining full coefficient interpretability inside the production decision engine.
]]>Authors: Kasidit Kokkhunthod Talit Jumphoo Wongsathon Pathonsuwan Atcharawan Rattanasak Rattikan Nualsri Sittinon Thanonklang Peerapong Uthansakul Monthippa Uthansakul
Wearable vibrotactile devices are increasingly used in virtual reality, teleoperation and neurorehabilitation, but objective EEG evaluation of glove-mediated touch remains limited. We compared EEG recorded during natural object interaction with EEG recorded when tactile feedback was reproduced through a vibrotactile smart glove. Data were collected with an eight-channel wireless headset while participants interacted with three object types (bottle, cube, and sphere) in natural-touch and glove-mediated conditions. An exploratory model trained on natural-touch data and tested on glove-mediated trials yielded rounded cross-condition accuracies of 83%, 78%, and 68% for bottle vs. rest, cube vs. rest, and sphere vs. rest, respectively. These findings suggest that some object-related EEG patterns may carry across conditions, but they should not be interpreted as evidence of physiological equivalence. Supplementary analyses using repeated-run evaluation, band-power and ERD/ERS summaries, temporal-window inspection, and channel ablation were included as cautious interpretability checks. The results underscore the need for larger subject-independent studies, stronger artifact-control pipelines, formal statistical testing, and richer haptic conditions before asserting equivalence to natural touch.
]]>Authors: Hoda Helmy Chaima Ben Rabah Ahmed Serag
Recent advances in artificial intelligence have significantly improved performance in medical imaging tasks such as segmentation, quantification, and report generation. However, most existing solutions operate as static pipelines with limited adaptability, quality assurance, and workflow-level reasoning. In this work, we present ARGUS, an agentic framework for multimodal medical image analysis that coordinates specialized agents within a unified architecture. An Orchestrator Agent interprets user requests, identifies the imaging modality, and assembles task-specific execution plans by selectively engaging processing, quantification, verification, knowledge retrieval, and reporting agents. This enables context-aware decision-making and dynamic workflow reconfiguration based on intermediate findings and runtime conditions. A key feature of ARGUS is its ability to supervise and contextualize analytical processes. The Verification Agent performs quality control by assessing intermediate artifacts against task-specific criteria, while the Knowledge Retrieval Agent enriches quantitative findings with evidence from the biomedical literature and established physiological reference ranges. Together, these components promote transparency, support automated error detection, and reduce the risk of propagating unreliable information through downstream stages. The framework was evaluated across three imaging domains: radiology (MRI), pathology (hematopathology), and ophthalmology (OCT). Quantitative evaluation demonstrated strong agreement between ARGUS and reference standards across pathology, OCT, and MRI tasks, achieving a cell-counting bias of −0.182 cells (MAE = 0.727), a full retinal thickness bias of −31.30μm (MAE = 37.63 μm), and MRI volumetric errors below 3 mL, while also achieving closer agreement with reference measurements than the evaluated general-purpose and domain-specific baseline systems. These results demonstrate the feasibility and potential value of agent-based orchestration for enabling adaptive, validated, and interpretable multimodal imaging workflows while providing a scalable foundation for complex multi-step clinical analysis.
]]>Authors: Miguel-Ángel Guillén-Ramos Héctor-Ricardo Hernández-de-León José-Armando Fragoso-Mandujano Elías Neftalí Escobar-Gómez Yair González-Baldizón Esvan-Jesús Pérez-Pérez
Wildfires are events in which fire spreads uncontrollably, destroying natural ecosystems and causing severe biodiversity loss. Driven by climate change, the frequency and intensity of these events have escalated, making effective risk mitigation a critical global priority. This study proposes a hybrid methodology for the identification and classification of wildfire risk zones, based on the combination of a convolutional neural network (CNN) with U-Net architecture and ResNet-50 backbone for semantic segmentation, together with a fuzzy inference system. The CNN processes high-resolution georeferenced RGB imagery to identify environmental patterns such as vegetation density and combustible organic matter. Evaluated through cross-validation, the CNN achieved a global IoU of 86.73% and a global F1-score of 92.82%, generating a risk classification into low, medium, and high levels. Subsequently, this categorical output is integrated into a fuzzy inference system along with meteorological variables (temperature, humidity, and wind speed). The fuzzy inference system dynamically adjusts the initial risk assessment according to meteorological conditions and generates an updated risk classification for the processed images. This approach significantly improves wildfire risk assessment, providing a data-driven tool for environmental management and early disaster prevention.
]]>Authors: Eva Brenner Luka Bulić Vilena Vrbanović Mijatović
Background: Postoperative kidney dysfunction is a common and serious complication in surgical patients. Kidney function is typically assessed using the estimated glomerular filtration rate (eGFR), most often calculated with the CKD-EPI equation based on serum creatinine. While several machine learning models have been developed to predict acute kidney injury, few have focused on predicting postoperative eGFR category. This study aimed to develop a machine learning model capable of classifying surgical patients into eGFR categories G1, G2, or G3+ in the early postoperative period, based on preoperative and intraoperative data. Methods: We developed the two-layer “eGFR-AI” architecture. In the first layer, two XGBoost models compute the probability of eGFR being above 89 and 59 mL/min/1.73 m2, respectively, and their outputs are passed to a second-layer logistic regression model that produces the final classification. The dataset included 200 patients admitted postoperatively to the intensive care unit of a tertiary academic hospital during the first half of 2024. Input features comprised age, sex, body mass index, type and duration of surgery, ASA status, presence of sepsis or shock at admission, and history of arterial hypertension, diabetes mellitus, or chronic kidney disease. Model performance was evaluated using accuracy, F1 score, and area under the ROC curve (ROC-AUC) on a held-out testing set. Feature importance analysis and statistical testing of associations with acute kidney injury were also performed. Results: On a held-out test set, the final model achieved an accuracy of 0.75, a weighted F1 score of 0.75, and a weighted ROC-AUC of 0.92 (balanced accuracy 0.76; Matthews correlation coefficient 0.65; Cohen’s κ 0.62). The first-layer models reached ROC-AUC values of 0.85 (eGFR > 89) and 0.96 (eGFR > 59). In a head-to-head comparison on the same partition, the stacked model performed comparably to standard baseline classifiers (multinomial logistic regression, random forest, support-vector machine, single multiclass XGBoost, CatBoost, LightGBM) without demonstrating superiority. Chronic kidney disease and presence of sepsis or shock at admission emerged as the strongest predictors. In an exploratory analysis (n = 8 events), all patients diagnosed with acute kidney injury fell into the G3+ category. Conclusions: In this single-center pilot study, “eGFR-AI” shows that early postoperative kidney function category can be predicted from a small set of routinely available preoperative and intraoperative variables, with performance comparable to standard classifiers and the added benefit of calibrated, interpretable category-level probabilities. Given the limited unicentric cohort and reduced category granularity, these findings should be regarded as preliminary and hypothesis-generating: they support the feasibility of the approach and motivate external, multicenter validation before any clinical application.
]]>Authors: George Melville Julian Yeomans
Previous research has shown sector-conditional asymmetry in implied volatility levels and in option returns. However, no prior work has parameterised that asymmetry at the effective-theta layer in a form that fires a non-discretionary rule trigger. This study supplies the parameterisation, its formulation, the first observation, and the data evidence. An effective theta is defined as Θe=αs,r⋅ΘBS, where ΘBS is the standard Black–Scholes (BS) theta and αs,r is a sector- and regime-conditional scaling factor. A SIMDEC decomposition is used to filter the input space and to determine the corner where α matters most. The framework is a bounded retrieval-and-deterministic compute system. The instruments are retrieved from cached market data and the learned layer’s outputs are constrained to that admissible set. Therefore, by construction, it cannot confabulate a fictitious or out-of-bounds instrument and the generative-class hallucination failure mode cannot occur. This concerns the groundedness and bounds of every output and is distinct from the accuracy of the regime and quality labels. SIMDEC supplies the joint-state filtering partition and, together with the Sobol variance decomposition, an explainability and attribution layer in which every position-level evaluation maps to an interpretable joint-state bin and a variance-share attribution. A “first observation” arising from a three-position long-call cohort traversing terminal decay is deployed using eight intraday states tracked on the trajectory at primary-source resolution and illustrates the relationship of the α parameterisation to existing market conditions. To examine the effectiveness of the approach, a SIMDEC dataset from the same deployment supplies population-level support across 12 sectors and a three-tier quality stratification. The dataset is the output of the THETA AI/ML pipeline—a multi-architecture deep-learning inference system that treats SIMDEC joint-state partitioning and Sobol variance decomposition as complementary interpretability inputs, with the regime classifier carrying the labels and the composite quality scorer carrying the stratification. The PC-based, token-free analytical procedure for regulated decision-making settings, together with an illustrative example of the asymmetry in the effective-theta provide a “next level” contribution to traditional option methodology.
]]>Authors: Hongbo Qiu Chenxuan Zhang Peixiao Fan Yuxin Wen Qianyi Yang
Background: With the deepening coupling of multi-microgrids (MMGs) and transportation systems in smart cities, maintaining frequency stability under extreme conditions increasingly relies on vehicle-to-grid (V2G) flexibility. However, existing V2G dispatch strategies often overlook the noticeable battery degradation caused by high-frequency regulation and the vulnerability of extensive communication networks to false data injection attacks (FDIAs), while the high-dimensional coordination of EV routing and discharging makes classical algorithms struggle to converge. Methods: To address these challenges, this study proposes a quantum-empowered degradation-aware V2G coordination framework for smart-city MMGs considering communication security and user travel demands. At the physical layer, an equivalent RC circuit-based battery degradation model and a traffic flow model are established to quantify capacity loss and travel delays. At the cyber layer, quantum key distribution (QKD) ensures unconditionally secure communication, while a quantum reinforcement learning (QRL) algorithm is developed to achieve fast convergence in high-dimensional multi-objective optimization. Results: Simulation results demonstrate that the proposed framework completely immunizes the system against FDIAs, effectively suppresses frequency fluctuations, and significantly reduces battery degradation costs while preserving user mobility. Conclusions: This framework provides a highly secure and user-friendly pathway for resilient smart-city frequency regulation.
]]>Authors: Hanjui Chang Dekai Kang Linrong Li Xin Yang Fei Long Jiaquan Li Rui Zhu Junhao Ye
In-Mold Electronics (IMEs) present a highly promising monolithic integration strategy for manufacturing miniaturized 3D MEMS optical tweezers, offering exceptional environmental adaptability and structural compactness. However, the precision of such optical systems is heavily constrained by the injection molding process. During the molding phase, high-pressure melt scouring and severe thermo-mechanical coupling frequently induce geometric misalignment, manifesting as node displacement, localized warpage, and residual stress accumulation in the embedded circuits. This displacement critically alters the cross-sectional area of conductive traces, leading to resistance fluctuations that can destabilize the driving current. According to American Wire Gauge (AWG) standards, ensuring the geometric fidelity of this sensor-CPU interconnect pathway is fundamental to maintaining signal integrity. To address these manufacturing bottlenecks, this study systematically investigates the process stability of IME circuits Cyclic Olefin Copolymer (COC) is strategically selected as the substrate material over Polycarbonate (PC) and Liquid Silicone Rubber (LSR) due to its ultra-high light transmittance, extremely low water absorption, and superior thermomechanical stability. Based on finite element simulation, a data-driven intelligent optimization framework is developed. Latin Hypercube Sampling (LHS) is first utilized to efficiently sample the multi-dimensional process space, comprising melt temperature, packing pressure, and packing time. To handle the non-stationary nature of process feedback signals, wavelet analysis is introduced to decouple high-frequency noise, extracting Wavelet Energy Entropy (WEE) as a highly robust dynamic metric for process stability. Subsequently, a hybrid NSGA-II-MOPSO multi-objective algorithm is deployed to cooperatively optimize the injection parameters. The simulation-based optimization results demonstrate a substantial enhancement in manufacturing precision. Under the optimal parameter configuration, the average node displacement of the embedded circuits decreases significantly from 0.034 mm to 0.014 mm, achieving a 58.82% reduction. Simultaneously, volumetric shrinkage drops from 5.755% to 4.832% (a 16.04% reduction), while residual stress is maintained well within the structural safety threshold of optical-grade polymers. By clarifying the deformation control mechanism during the manufacturing phase, this study provides a highly reliable, data-driven methodological framework for the precision mass production of micro-nano optical systems.
]]>Authors: Mlungisi Ntombela
Transmission power losses significantly affect the efficiency, reliability, and economic operation of modern electrical power systems. This study proposes a Hybrid Genetic Algorithm–Particle Swarm Optimization (HGAPSO) framework for transmission loss minimization in the IEEE 118-bus power system. The proposed approach combines the global exploration capability of Genetic Algorithms (GAs) with the rapid convergence characteristics of Particle Swarm Optimization (PSO) to optimize generator voltage settings, transformer tap positions, and reactive power compensation while satisfying all operational constraints. The HGAPSO framework was developed and implemented in MATLAB R2024a and evaluated using the IEEE 118-bus test system. The simulation results demonstrate that the proposed method reduced transmission losses from 132.8 MW under the base-case condition to 98.6 MW, representing a 25.75% reduction in total network losses. In addition, the optimized operating conditions improved the minimum bus voltage from 0.914 p.u. to 0.972 p.u., while the average voltage deviation decreased from 0.062 p.u. to 0.019 p.u. These voltage profile improvements were achieved as secondary benefits of the transmission loss minimization process and the enforcement of system operating constraints. Furthermore, the HGAPSO algorithm exhibited superior convergence performance, reaching the optimal solution within 82 iterations compared to 185 iterations for GA and 124 iterations for PSO. The results confirm that the proposed HGAPSO framework provides effective transmission loss reduction, faster convergence, and reliable network operation compared with standalone optimization techniques. The proposed methodology offers a robust and computationally efficient solution for large-scale power system optimization, optimal power flow studies, and smart grid applications.
]]>Authors: Yryskeldi Siddi Giorgio Delzanno Daniele D’Agostino
The increasing demand for real-time, energy-efficient machine learning (ML) models in edge and embedded scenarios highlights the limitations of traditional CPU-based inference pipelines. This work presents a pipeline that integrates Automated Machine Learning (AutoML) with High-Level Synthesis (HLS) to enable efficient deployment of ML predictors on Field-Programmable Gate Arrays (FPGAs). Using Auto-Weka for automated algorithm selection and Bayesian hyperparameter optimization, followed by model reimplementation and parameter extraction using scikit-learn and synthesis through the AMD/Xilinx Vitis HLS toolchain, the proposed workflow combines data-driven model exploration with hardware-oriented implementation. The pipeline adopts a hybrid, human-in-the-loop approach, reflecting current practical constraints in bridging heterogeneous software and hardware environments. Experimental results show up to a 9% latency reduction compared to CPU-based inference, more than a 7× throughput improvement through parallel FPGA instantiation, and more than an order of magnitude improvement in energy efficiency when evaluated in terms of energy per inference. The proposed methodology provides a reproducible and extensible workflow for integrating AutoML-based model discovery with FPGA deployment, while highlighting both the benefits of hardware acceleration and the remaining challenges in AutoML–hardware integration.
]]>Authors: Koffka Khan Winston Elibox Shanta Ramnath
Federated learning (FL) can support environmental monitoring when facilities cannot pool raw operational records, but ordinary FL does not encode permit direction, regulatory margins, censoring, missingness, or facility-specific compliance regimes. This study introduces Federated Residual-Element Optimization (FREO), a compliance-aware environmental artificial intelligence (AI) framework that converts effluent measurements into signed, unit-consistent regulatory residuals and communicates compact residual elements rather than raw samples. Ten de-identified Environmental Management Authority effluent workbooks, treated as workbook-level facility-category clients, produced 4036 assessable observations, 321 facility-month states, and 311 next-period prediction records. A boundary-purged temporal split removed train/test target-input overlap, leaving 204 training and 97 held-out records. On the within-facility next-period high-burden task, FREO achieved AUROC 0.966, AUPRC 0.923, F1 0.852, and Brier score 0.079, with facility-month bootstrap intervals of [0.924, 0.996], [0.810, 0.990], [0.727, 0.941], and [0.056, 0.104], respectively. Matched calibrated baselines indicate that residual features and training-only facility-prior calibration explain much of the gain; FedAvg-Cal slightly exceeded FREO on AUPRC, F1, and Brier. FREO is therefore positioned as an auditable residual-element workflow for raw-data-local regulatory screening, not as a universally superior optimizer.
]]>Authors: Tihomir Orehovački
Large language models (LLMs) are changing how students approach problem solving, code interpretation, and software development. Successful task completion with AI assistance, however, does not necessarily indicate conceptual understanding of underlying programming principles, making it difficult to determine how programming skills develop over time. This study examines whether students’ critical engagement, collaborative learning practices, and exploratory use of LLMs are associated with self-reported programming competence, coding practices, and longer-term knowledge retention. Survey data from 189 students with varying levels of LLM use in educational and coding-related contexts were analyzed using partial least squares structural equation modeling (PLS-SEM). The findings suggest that students perceive LLMs as more educationally valuable when they actively question, reinterpret, and adapt AI-generated responses rather than accept them as final answers. Students who interacted more reflectively with AI outputs reported stronger perceived competence and more deliberate attention to code organization, readability, and maintainability. Collaborative use corresponded to broader development of programming abilities, whereas creative experimentation was more closely related to stylistic refinement and perceived benefits for longer-term retention. Active engagement with AI-generated material may therefore promote analytical reasoning and deeper conceptual involvement instead of merely accelerating code production. By moving beyond technology adoption and productivity-oriented perspectives, the study highlights the role of learner agency in shaping LLM-supported programming education.
]]>Authors: Miguel Angel Camargo-Rojas Gabriel Sánchez-Pérez José Portillo-Portillo Linda Karina Toscano-Medina Aldo Hernández-Suárez Jesús Olivares-Mercado Héctor Manuel Pérez-Meana
The automated digitization of handwritten electoral results is critical for ensuring transparency and speed in democratic processes. While recurrent sequence-to-sequence models (e.g., CRNN+CTC) achieve high accuracy, they inherently violate the strict latency constraints of high-throughput administrative environments. Conversely, standard lightweight CNNs exhibit suboptimal performance on the long-tail distribution of high-cardinality scenarios. To bridge this gap, this study reformulates sequence recognition into a latency-bound classification task. We propose a specialized Handwritten Digit Sequence Recognition (HDSR) framework for the Mexican Preliminary Election Results Program (PREP) based on a modified ResNet-18 architecture. The methodology introduces an asymmetric stride designed explicitly to preserve the 1:3 horizontal feature resolution of electoral tally sheets, integrating a lightweight Convolutional Block Attention Module (CBAM) in deep stages to refine classification across 1001 possible sequences. Leveraging a megadiverse dataset of 3.77 million real-world images, the model was trained using AdamW and label smoothing to mitigate human-induced label noise. Results demonstrate a global accuracy of 97.82% and a significant improvement in Macro-Precision (0.8878) for rare sequences. With an inference latency of 9.1 ms on standard CPU hardware, the proposed solution offers a scalable, high-confidence alternative that prioritizes spatial preservation and fail-controlled deployment.
]]>Authors: Sanjay Mishra Ganesh R. Naik
When a large language model (LLM) answers a question using retrieved documents, retrieval-augmented generation (RAG) is the standard approach. Retrieving more documents improves answer accuracy but increases cost and response time; retrieving fewer documents saves resources but may miss critical information. Most existing RAG systems sidestep this dilemma by applying the same retrieval setting to every query, regardless of how simple or complex the question is. This wastes budget allocation on easy questions and under-serves hard ones. This paper introduces Cost-Aware RAG (CA-RAG), a routing framework that solves this problem by treating each query individually. For every incoming question, CA-RAG selects the most suitable retrieval strategy from a fixed menu of four options, ranging from no retrieval to fetching the top k=10 most-relevant documents. The selection is driven by a scoring formula that balances expected answer quality against predicted cost and response time. The weights in this formula act as dials: adjusting them shifts the system toward speed, savings, or quality without any retraining. CA-RAG is built on Facebook AI Similarity Search (FAISS) for document retrieval, OpenAI gpt-4o-mini for generation, and text-embedding-3-small for dense retrieval embeddings. We evaluate CA-RAG on a benchmark of 28 queries. The router assigns different strategies to different queries, achieving 26% fewer billed tokens compared to always using heavy retrieval and 34% lower response time compared to always answering without retrieval, while maintaining answer-quality parity in both cases. Further analysis shows that most savings come from simpler queries, where heavy retrieval was unnecessary. All results are reproducible from logged comma-separated value (CSV) files. CA-RAG demonstrates that a small but well-designed set of retrieval strategies combined with lightweight per-query routing can meaningfully reduce the cost and latency of LLM deployments without compromising answer quality.
]]>Authors: George Dueñas Sergio Jimenez Geral Eduardo Mateus Ferro
Large Language Models (LLMs) achieve strong performance on standardized examinations, yet the language-mediated mechanisms through which they assess answer options in Multiple-Choice Questions (MCQs) and diverge from human difficulty judgments remain poorly understood. We argue that predicting the difficulty of MCQs provides a lens for studying how LLMs represent plausibility, uncertainty, and error across competing response options. Although recent deep machine learning approaches achieve competitive accuracy through large feature sets and complex architectures, their limited interpretability reduces their value for understanding model behavior. We propose an interpretable framework that decomposes item difficulty into LLM-based plausibility estimates over response options. These estimates are elicited through direct prompting and pairwise contrastive comparisons, and then integrated into a rational model that expresses item difficulty as a ratio between the plausibility of distractors and the plausibility of the correct option. We evaluated this approach on two high-stakes datasets. Using the United States Medical Licensing Examination (USMLE) dataset, the model achieved a Root Mean Squared Error (RMSE) of 0.277, comparable to previous approaches, while reducing the representation of the underlying elements from hundreds of features to only three parameters. Under Spearman rank correlation, the model reached ρ=0.427 on USMLE, representing a 70.8% relative improvement over previously reported results, and ρ=0.488 on ICFES, a new dataset. A complementary ranking analysis further reveals a systematic inversion between LLM-based difficulty judgments and those of experts, exposing divergences between model-internal assessments and human response patterns. These findings position option plausibility based on divide-and-conquer prompting as a principled framework for probing LLM decision processes, their rank-order misalignment with human response patterns, and their challenges in educational settings.
]]>Authors: Gilberto Huesca Yolanda Martinez-Trevino Claudia Jiménez González David Cantú Delgado Christelle Navarrete Antonio Cedillo-Hernandez Ricardo Quintero Meza
The rapid integration of Generative Artificial Intelligence (GenAI) into programming education has raised important questions regarding its impact on learning processes, conceptual understanding, and technological dependency. This study analyzed the effects of four GenAI-supported instructional strategies in an introductory programming course for undergraduate engineering students. A multi-group quasi-experimental pre-test–post-test design was implemented involving 686 students distributed across 53 class groups, from 10 campuses, taught by 32 professors. The instructional conditions included Quizzes for Self-Regulation, Github-Copilot-assisted learning, Prompt Problems with Iterative Refinement, and Flipped Learning enhanced with GenAI, which were compared against a traditional teaching approach. Learning outcomes were measured using normalized learning gain, while statistical analyses were conducted using non-parametric methods due to deviations from normality and heteroscedasticity. Results indicate that GenAI integration did not produce statistically significant overall differences in learning gain when all GenAI-supported strategies were analyzed as a single cluster compared to traditional instruction. However, differences emerged between specific strategies, with Quizzes and Copilot-based approaches having higher median learning gains than Prompt Problems and Flipped Learning strategies. No statistically significant differences associated with gender were identified. These findings suggest that the effectiveness of GenAI in programming education depends less on the mere presence of the technology and more on the pedagogical conditions under which it is integrated into the teaching–learning process.
]]>Authors: Mishell Cadena-Yanez Angela Bernardini Marisol Gómez
This study examined how healthcare professionals (HCP) perceive and trust AI in clinical settings, and what kinds of explanations they need to use it effectively. Using interviews with six HCPsand a questionnaire (N=41), the findings show that trust was higher for administrative tasks and lower for direct clinical decisions, regardless of their prior AI or clinical experience or AI literacy. Clinical validation and algorithmic bias ranked above explainability as trust factors, indicating that explainability is not a primary trust-building mechanism. However, HCP consistently demanded explanations as a tool to support their own clinical reasoning, preferring to receive them across all clinical contexts rather than only under disagreement with the system, and valued grounding in medical evidence and consistency with clinical protocols over clarity or simplicity. These findings argue for an HCXAI design approach that treats explainability as a critical-reasoning tool rather than a primary trust mechanism.
]]>Authors: Yildiray Kabak Gokce B. Laleci Erturkmen Mert Gencturk Tuncay Namli A. Anil Sinaci Ruben Alcantud Corcoles Cristina Gómez Ballesteros Pedro Abizanda Volkan Atmis Asuman Dogac
Background: Evidence-based clinical guidelines are essential for high-quality care yet translating them into personalized clinical decision support remains resource-intensive and time-consuming. Large language models (LLMs) show promise for supporting clinical decision-making, but their limited access to patient-specific data and explicit guideline sources constrains trustworthiness, personalization, and clinical applicability. Retrieval-augmented generation (RAG) addresses part of this challenge by grounding model outputs in curated evidence sources; however, true personalization requires structured access to electronic health record data. Methods: This study presents FHIR-RAG-MEDS, a medical decision support system that integrates HL7 Fast Healthcare Interoperability Resources (FHIR) with an RAG-enhanced LLM to enable patient-specific, guideline-concordant clinical recommendations. Through SMART on FHIR, the system retrieves real-time patient data from FHIR servers, generates structured medical summaries, and incorporates this personalized context into the RAG pipeline, grounding responses in evidence-based clinical guidelines stored in a vector database. Results: FHIR-RAG-MEDS was evaluated using 139 physician-generated clinical questions covering dementia, chronic obstructive pulmonary disease, hypertension, and sarcopenia. Performance was assessed using automated metrics, RAG-specific evaluation frameworks, and independent expert physician review. The system consistently outperformed state-of-the-art medical LLMs, demonstrating higher semantic accuracy, improved faithfulness to guideline content, and stronger clinical relevance. Conclusions: Integrating HL7 FHIR with RAG-based LLMs enables trustworthy, personalized clinical decision support, bridging the gap between static language models and real-world, patient-centered care.
]]>Authors: Maryam Khanian Najafabadi Bushra Naeem Touraj Khodadadi Saman Shojae Chaeikar Zawar Shah
This study proposes a time series-aware Large Language Model (LLM)-driven feature engineering framework for tabular prediction tasks. Existing automated feature engineering methods, including LLM-based approaches such as CAAFE and OCT-Tree and established libraries such as tsfresh and Featuretools, can generate useful features for general tabular data but do not explicitly address temporal availability constraints in time series settings. This can lead to data leakage when variables that are only available after the prediction event are used directly during model training. To address this limitation, the proposed framework classifies variables into antecedent features, consequent features, and historical aggregated features. The key innovation is that consequent variables are not discarded to prevent leakage but are instead routed into a leakage-safe historical aggregation pipeline, recovering predictive signal from post-event variables through temporally valid past values. The framework guides an LLM to generate structured feature engineering configurations, applies temporally valid transformations, performs feature selection, and evaluates the selected features using predictive models. A formal leakage control mechanism ensures that all aggregations use strictly past observations, applied within entity groups and before the temporal train–validation–test split. The framework is evaluated on two time series tabular tasks: Tesla stock prediction and English Premier League match outcome prediction. The results show that the proposed approach improves predictive performance compared with raw-feature baselines and selected existing automated feature engineering methods. On the Tesla dataset, the framework reduced MAE compared with both the baseline and the reported OCT-Tree result. On the EPL dataset, it improved accuracy compared with the odds-only baseline and the reported Azure ML preprocessing result. These findings suggest that combining LLM reasoning with explicit temporal constraints is a practical direction for automated feature engineering in time series tabular machine learning.
]]>Authors: Meryem Chajia El Habib Nfaoui Soufiyan Ouali
Customer churn has become a major challenge in the retail industry, where customer loyalty directly affects business success and sustainability. Despite the significant progress in Artificial Intelligence, especially in prediction tasks, its use in the retail e-commerce domain remains limited and underexplored; this is due to the scarcity and limited quality of available datasets. To address these challenges, this paper proposes a churn prediction approach designed to handle data scarcity while ensuring accurate performance. We experimented with a combination of various feature selection techniques along with several Machine Learning and Deep Learning models to evaluate their performance on a limited tabular dataset. The impact of feature selection on predictive performance was also systematically analyzed. The results demonstrated that feature selection plays an important role in improving model performance by identifying the key features that have the most significance to the classification task. The analysis showed that the L1-based Logistic Regression feature selection method combined with the Extreme Gradient Boosting classifier achieved the best performance, with a Macro F1-score of 95.25%. Based on these results, companies can identify potential churners and implement retention strategies. These findings may provide a useful reference point for future researchers in the retail e-commerce industry.
]]>Authors: Ashley Allen Alexios Mylonas Stilianos Vidalis Nikolaos Pitropakis
The last five years have seen considerable growth in the topic of peer-to-peer (P2P) federated learning (FL). This framework removes the central coordinating server used in conventional federated learning and instead requires participating nodes to manage model training, peer selection, communication, aggregation, and trust directly. This provides a promising route for privacy-preserving and decentralised machine learning, but it also introduces unresolved challenges in topology selection, participant incentivisation, communication efficiency, security, and evaluation. Existing studies frequently evaluate proposed methods under narrow assumptions, such as static network membership, homogeneous devices, fixed bandwidth, limited topology choices, and public benchmark datasets. Existing surveys also tend to present taxonomies of decentralised federated learning rather than synthesising how topology, incentives, and communication algorithms jointly affect deployability. This paper reviews recent work on peer-to-peer federated learning across three connected dimensions: network topology, incentive mechanisms, and communication algorithms. We compare the topologies, datasets, experimental assumptions, incentive designs, communication strategies, and open issues reported in the literature. The review shows that highly connected topologies tend to improve convergence but increase communication overhead and vulnerability to bottlenecks; sparse and dynamic topologies improve efficiency but create challenges for convergence, reliability, and node drop-out. Incentive mechanisms increasingly combine reward, reputation, validation, and punishment but remain weakly validated under realistic churn, heterogeneous resources, and adversarial behaviour. Communication algorithms reduce bandwidth through gossip, sparsification, prediction, routing, and multi-step aggregation but often trade communication savings against accuracy, robustness, and generalisability. Across all three areas, the field lacks standardised benchmarks, reproducible experimental settings, and realistic evaluation under unstable peer-to-peer conditions. We conclude by identifying cross-cutting research gaps and recommending future work on dynamic topologies, heterogeneous devices, real-world datasets, incentive robustness, and comparable benchmarking.
]]>Authors: Mohammad Odeh Ahmad Hasasneh
Precise weather forecasting plays a critical role in sectors such as agriculture, transport, energy management, and climate change adaptation, and machine learning and deep learning algorithms have been widely used for data-driven time series forecasting problems. In this work, we explore the application of machine learning and deep learning models for multi-weather variable forecasting in a dataset recorded over a period of ten years (2015–2025) for five weather stations in Palestine. The dataset comprises measurements for temperature, relative humidity, wind speed, precipitation, atmospheric pressure, and sunshine hours. To avoid the issue of temporal leakage, a chronological training, validation, and test set splitting approach was used in the evaluation experiments. The models used in this study include ARIMA, SARIMA, Random Forest, XGBoost, CNN, LSTM, GRU, ConvLSTM, CNN-GRU, and CNN-LSTM with station embeddings. Our experimental results indicate that the XGBoost model achieved the highest performance in predicting temperature and relative humidity (R2 = 0.953 and R2 = 0.670, respectively), while deep learning methods exhibited high accuracy across several weather features. The CNN-LSTM model was successfully able to learn temporal–spatial patterns via station embeddings, while recurrent neural networks performed impressively in forecasting sunshine hours and atmospheric pressure.
]]>Authors: Yunfei Wang Weidong Jia Mingxiong Ou Xiang Dong Shiqun Dai Rong Zhang Yaning Wang Wenrui Zhu
Real-time kinematic (RTK) navigation, which enables centimeter-level positioning accuracy through carrier-phase differential correction, provides high-accuracy positioning for orchard robots, but short-term outages caused by canopy occlusion and signal interference may interrupt path guidance and increase lateral drift. To address this issue, this study proposes a vision-based takeover navigation method for orchard robots under short-term RTK failure conditions. First, an improved YOLOv11-based road segmentation and completion model, termed YOLOv11-VF, was developed. By introducing a Squeeze-and-Excitation (SE) channel attention mechanism, the model jointly perceives visible road regions and occluded road completion regions, thereby producing continuous and complete road semantic representations. Second, a structured geometric road representation was constructed from the segmentation results to extract the navigation reference line, and a joint direction-position constraint mechanism was established by integrating the reference line with the robot reference view axis. A hierarchical constraint strategy based on a travel corridor and a deadband region was further designed to jointly determine heading deviation and lateral drift. Finally, road segmentation, navigation-line extraction, parameter analysis, and vision-based takeover experiments were conducted in a standardized orchard environment. The results showed that YOLOv11-VF achieved Precision, Recall, AP50, mAP@0.5:0.95, and F1 values of 92.31%, 88.56%, 94.40%, 67.41%, and 90.40, respectively, showing the best overall segmentation performance among all compared models while maintaining good real-time performance. The proposed method also demonstrated high consistency in navigation-line extraction and maintained mean absolute deviations of 0.0176 ± 0.0041 m to 0.0718 ± 0.0138 m during RTK outage intervals over 10 repeated trials, indicating good path-following capability and operational stability.
]]>Authors: Lujain Abdulrahman Alawwad Mohamed El Bachir Menai
While aspect-based sentiment analysis (ABSA) has made significant progress in the identification of explicit opinion targets, the more challenging case of implicit aspects remains insufficiently studied. Implicit aspect extraction is particularly challenging, as it relies on contextual and semantic cues and requires systems to infer what reviewers mean rather than what they state explicitly. A four-component hybrid pipeline is proposed for explicit and implicit aspect extraction, formulating the task as controlled text generation. The pipeline combines (i) a fine-tuned decoder-only large language model as a generative baseline, (ii) an iterative residual generation strategy that recovers multiple aspects through successive masked generation passes, (iii) paraphrase-based input transformation to broaden the contextual signal, and (iv) domain-specific knowledge graphs activated by linguistic signals to infer implicit aspects. The novelty lies not in the individual components themselves but in their principled orchestration and the linguistically motivated gating logic governing the activation of each stage. Extensive experiments are conducted on eight benchmark ABSA datasets spanning both English and Arabic: SemEval-2014, SemEval-2015, SemEval-2016, ACOS, and M-ABSA for English; and SemEval-2016, HAAD, and M-ABSA for Arabic. The proposed solution outperforms strong baseline methods and recent state-of-the-art models on English datasets, with F1-scores of 0.8533, 0.713, 0.7859, 0.793, and 0.664, respectively. On Arabic datasets, the best-performing configurations achieve F1-scores of 0.7632, 0.4765, and 0.7656 on SemEval-2016, HAAD, and M-ABSA, respectively, with the knowledge-graph component providing consistent and statistically significant gains for implicit aspect identification in both languages. These results demonstrate the effectiveness of generative modeling, iterative generation, paraphrasing, and structured knowledge for aspect extraction and highlight the potential of the proposed approach for implicit aspect identification, in particular for morphologically rich languages such as Arabic, where annotated resources remain scarce.
]]>Authors: James Vakilian Fareed Ud Din Edmund J. Sadgrove Mohammadreza Haghighat Niusha Shafiabady
The integration of international students into host countries and their cultures is a multifaceted challenge that significantly impacts their academic success and well-being. This study leverages Explainable Artificial Intelligence (XAI) to model and interpret variables associated with the self-rated integration of 175 international students at Charles Darwin University (CDU) in Australia, using data from a 42-question survey. Employing machine learning models, including Decision Tree (DT) and Gradient Boosting Machine (GBM), we use XAI techniques to identify variables most strongly associated with students’ self-rated integration, including career confidence, perceived future happiness, and perceived career obstacles. SHAP analyses and partial dependence plots provide global and instance-level insights, revealing both the magnitude and directional effects of these features. The findings highlight the predictive relevance of psychological and social variables in students’ self-rated integration, offering exploratory insights that inform targeted support programs. By enhancing model transparency through XAI, this research fosters trust in AI-driven educational interventions, addressing ethical considerations and promoting equitable outcomes for diverse student populations.
]]>Authors: María Isabel A. Benítez Serafín Moral-García Joaquín Abellán
Artificial intelligence systems that learn or reason from finite empirical data often require uncertainty representations that go beyond a single precise probability distribution. This is especially relevant when observations are scarce, incomplete or not reliable enough to support precise probabilistic assessments. In current data-driven AI tools, empirical information extracted from data must often be converted into a structured uncertainty model before it can be used for reasoning, learning or decision support. The singleton intervals induced by NPI-M and A-NPI-M provide such a representation, since they express the predictive information obtained from the observed data without introducing externally chosen cautiousness parameters. Evidence theory is useful in this context because it allows partial support to be assigned to sets of alternatives, making it suitable for representing imperfect knowledge in AI systems. This paper studies how Non-Parametric Predictive Inference for multinomial data (NPI-M) can be connected with evidence theory through reachable probability intervals. Since the exact NPI-M model does not directly define a credal set, we focus on its approximated version, A-NPI-M, which preserves the NPI-M singleton bounds and represents them through reachable probability intervals. We analyze whether the resulting credal set can be represented exactly by a belief function, showing that this is not possible in general, although exact representations may exist in particular cases. Motivated by this limitation, we construct a basic probability assignment whose belief and plausibility values reproduce the A-NPI-M singleton bounds. The resulting belief function preserves the marginal interval information of A-NPI-M while adding an evidential structure on composite events, and its associated set of compatible probability distributions is included in the A-NPI-M credal set. The construction is presented by cases, illustrated with numerical examples and compared with the belief-function representation of the Imprecise Dirichlet Model. The proposed model provides a theoretical representation layer that may support uncertainty-aware AI procedures by transforming empirical predictive information into structured imperfect knowledge before reasoning, learning or decision-support criteria are applied.
]]>Authors: Igor Adamenko Orpaz Ben Aharon Yehudit Aperstein Alexander Apartsin
Urban environments are saturated with imaging sensors deployed for purposes unrelated to vehicle identification, from ATM and dashboard cameras to pole-mounted CCTV and smartphones. We term the use of such non-purpose-built sensors for secondary inference “opportunistic sensing”; its central question is where, under uncontrolled capture conditions, AI-enabled restoration remains reliable. This paper introduces recoverability maps, a task-agnostic methodology for quantifying that boundary, and applies it to oblique-view license plate recognition (LPR). It pairs a full-grid synthetic sweep of the degradation space with two summary measures: a boundary area-under-curve for coverage and a reliability score F for the frequency and depth of interior unrecovered pockets. For LPR, the space is the oblique-angle grid [0°,89°]2 sampled by Scrambled Sobol sequences, and the utility is plate-level optical character recognition (OCR) accuracy. Within this synthetic benchmark, approximately 90–92% of the angle grid is recoverable (best single model to union of restoration arms), recovery degrades sharply beyond roughly 80° in both axes, and lateral rotations are harder to reconstruct than elevational ones. Five restoration architectures cluster within a narrow AUC band of 0.89–0.93, and share the same α/β asymmetry, so the recoverable region is set primarily by sensing geometry, with architecture affecting efficiency and interior consistency; discriminative architectures outperform generative models. The methodology is validated on real plates: on CCPD and the Brazilian legacy and Mercosur layouts of RodoSol-ALPR, restoration raises held-out extreme-angle recognition by +15 to +38 exact-match points under plate-specialized recognizers, and the discriminative-over-generative ordering reproduces on real data.
]]>Authors: Leonardo Eleuterio Ariello Kelvin Wang David Newman-Toker Jee Bang David P. W. Rastall
Background: Artificial intelligence in medicine has largely relied on supervised training of disease-specific models, limiting scalability in conditions where labeled data are scarce. Large language models (LLMs), which encode broad medical knowledge through large-scale pretraining, offer an alternative paradigm in which structured physiological measurements can be interpreted directly without task-specific model training. Objective: To evaluate whether smartphone-derived ocular motor biomarkers can be translated into clinically meaningful probabilistic assessments of Huntington’s disease (HD) using general-purpose LLMs operating as inference engines. Methods: In this prospective proof-of-concept study, 26 participants (13 with genetically confirmed HD and 13 age-matched controls) completed a standardized ocular motor assessment using a custom smartphone application. Quantitative eye-movement metrics were validated against expert neurologist ratings. Structured physiological features were then provided to four general-purpose LLMs without task-specific training or diagnostic labels, and the models generated an AI-Assigned HD Probability Score (HAIPS). Discriminative performance and associations with clinical severity measures were evaluated. Results: Smartphone-derived ocular motor metrics showed strong agreement with clinician assessments (Spearman ρ = 0.76–0.95; all p < 0.001), confirming preservation of clinically meaningful physiological signals. LLM-derived HAIPS distinguished HD from controls with high accuracy (AUC 0.879–0.944), with no significant differences across models. Discrimination was statistically equivalent to a supervised logistic regression model trained on the same features. HAIPS correlated strongly with established measures of disease severity, including cognitive (MoCA, ρ = −0.86), functional (TFC, ρ = −0.74), and motor impairment (UHDRS, ρ = 0.85) (all p ≤ 0.003). Conclusions: Structured ocular motor biomarkers acquired using a consumer smartphone can be translated into clinically meaningful probabilistic assessments of HD by general-purpose LLMs without disease-specific model training. These findings support a framework in which physiologically grounded digital biomarkers are coupled with general-purpose inference models, potentially enabling scalable assessment in rare neurological diseases where labeled data are limited.
]]>Authors: Tim Geppert Andreas Block Maria Rothstein Mario Gellrich
Background: Artificial intelligence (AI) adoption is increasingly reported as a priority for organizations, yet they face a growing, fragmented body of evidence concerning the factors that influence successful AI integration. Method: To identify the relevant factors for organizational AI adoption, we conducted a systematic literature review (SLR) following PRISMA guidelines, which yielded 37 quantitative empirical studies. From these studies we extracted 1229 paper-item instances, of which 810 were retained after applying structured exclusion criteria to develop a domain model relevant to organizational AI adoption. The model’s content validity was assessed and supported through expert feedback using the Content Validity Index (CVI) methodology. Results: We organized 24 subclusters into nine main clusters across the three dimensions Technology (Enablers, Usability, Trust), Organization (Leadership, People, Process), and Environment (Market, Regulatory, Partner). Our analysis suggests that workforce skills, perceived intelligence, and resources are among the most frequently studied and positively associated antecedents of AI adoption, and that constructs related to AI explainability and control (human-in-the-loop oversight) have received little research attention and remain underrepresented despite growing regulatory requirements such as the EU AI Act. Conclusions: The resulting domain model provides an empirically grounded classification of organizational AI adoption factors and can serve as a foundation for future measurement instruments.
]]>Authors: Taehyeon Kim Kyung-Taek Lee
Group Relative Policy Optimization (GRPO) streamlines the alignment of Large Language Models (LLMs) and Vision–Language Models (VLMs) by eliminating the Critic model. However, its efficiency heavily depends on the group size, G. While a larger G improves reward estimation and stabilizes the Advantage, Ai, it drastically increases VRAM usage and reduces throughput. Standard heuristics like a fixed G of 64 create significant bottlenecks in resource-constrained settings. This paper introduces an Efficiency-Aware optimization framework utilizing Multi-fidelity Bayesian Optimization and Hyperband (BOHB) to dynamically identify the optimal group size, G∗. The method uses a multi-objective function that balances reward accuracy, Ai variance, and hardware utilization, applying z-score normalization. By employing Successive Halving to quickly evaluate candidates at low fidelity, the framework reduces search costs by up to 74% compared with random search. Tested across text-only LLMs (Qwen2.5-7B/1.5B) and multimodal VLMs (Qwen2.5-VL-3B), the framework demonstrates that the discovered G∗ saves up to 72.5% in VRAM compared with the baseline of 64, while maintaining reward accuracy within 5.8%. Sensitivity analyses on hyperparameters like λ, α, and β confirm the framework’s robustness. Rather than treating group size as a mere engineering heuristic, this study establishes a principled methodological advance by formalizing the trade-off between statistical estimation stability and hardware constraints into a unified optimization framework for resource-efficient RLHF.
]]>Authors: Rodrigo Garcia Horderlin Robles
Precision Livestock Farming (PLF) increasingly relies on edge-deployed sensors to monitor bovine behaviors, fostering improved welfare and management. However, behavioral data naturally expands over time and presents severe class imbalances due to animals’ predominantly sedentary routines. When continuous sequential updates are required without access to historical datasets, deep learning methods frequently succumb to catastrophic forgetting. This study introduces an ultra-lightweight (∼0.85 MB) Continual Learning (CL) architecture built upon a CNN-BiLSTM feature extractor, tailored to process multivariate Inertial Measurement Unit (IMU) streams. We exhaustively evaluated baseline Naïve Fine-Tuning against Elastic Weight Consolidation (EWC), Learning without Forgetting (LwF), and episodic Replay under three rigorous real-world paradigms: Class Incremental, Subject Incremental (domain shift), and Imbalanced Realistic scenarios. Our empirical findings expose the fragility of static paradigms: in Class Incremental expansions, Naïve Fine-Tuning collapsed to an Average Accuracy of 33.33%. Conversely, Experience Replay emerged as the most robust defense, achieving a statistically significant Average Accuracy of 74.64 ± 6.77% across multiple random seeds. Furthermore, LwF effectively mitigated structural variations across unseen animal domains (Subject Incremental) without requiring raw data buffers. Notably, under severe biological class imbalances (Imbalanced Cumulative), the architecture proved highly resilient, maintaining 98.46% Average Accuracy and retaining perfect minority class recall. This research validates the operational feasibility of deploying adaptive, privacy-preserving CL frameworks directly on low-power wearable devices for lifelong livestock monitoring.
]]>Authors: Muhammad Ali Shafique Imran Latif Hayat Ullah Alex C. Newkirk Arslan Munir
The rapid scaling of large language model (LLM) training has intensified demand for Graphics Processing Unit (GPU) clusters balancing throughput with energy efficiency. While NVIDIA’s H100 and B200 architectures are increasingly deployed in production datacenters, their comparative behavior under distributed training remains insufficiently characterized beyond vendor specifications, leaving datacenter operators without empirical guidance on metrics such as TFLOPs/kW and tokens-per-kilojoule. This work presents a system-level evaluation of single-node 8× H100 and 8× B200 configurations using Distributed Data Parallel (DDP) training across LLMs and vision–language models (VLMs) ranging from 7B to 32B parameters, spanning various real AI workload scenarios. We benchmark end-to-end throughput, utilization, power, energy, TFLOPs/kW, and tokens-per-kilojoule, complemented by architectural analysis explaining observed behavioral differences. Across LLM workloads, B200 achieves higher utilization (1–6%), faster training (up to 15%), and greater compute efficiency (up to 32% higher TFLOPs/GPU), attributable to higher memory bandwidth and large streaming multiprocessor (SM) count. However, B200 exhibits lower TFLOPs/kW and tokens-per-kilojoule, revealing a fundamental trade-off: throughput gains come at a measurable energy cost per useful token. VLM results further expose model-dependent asymmetries, with B200 consuming disproportionately more energy for lighter compute kernels due to elevated baseline power draw. These findings provide an empirical framework distinguishing compute efficiency from energy efficiency across next-generation GPU nodes, offering practical guidance for energy-aware AI datacenter design.
]]>Authors: Duc Nghia Vu Janos Demetrovics
The Q-Learning Relational Game (QLRG) framework provides a theoretically rigorous method for identifying minimal winning coalitions in online social networks (OSNs) under the restrictive assumption of global agent symmetry or uniform matroid structure. Real-world OSNs, however, exhibit significant asymmetry. This paper introduces the Cluster-Based Q-Learning Relational Game (C-QLRG), a practical extension that relaxes the global symmetry requirement by leveraging community structure. We partition the agent set into communities with bounded internal variation and represent the state solely by community membership counts of the seed set. Because the closure operator already captures all eventual influence spread, the problem reduces to a sequential seed selection task where the agent decides, at each step, from which community to add the next seed. We prove that the optimal Q-function of a suitably regularized reach-efficiency objective is Lipschitz continuous and derive a performance bound for the learned policy. The full algorithm is presented, and its complexity is analyzed. Empirical evaluations on a synthetic asymmetric network and Zachary’s Karate Club demonstrate that C-QLRG is highly sensitive to reward parameters, where default settings lead to premature stopping, but parameter tuning combined with a corrected minimality verification recovers high-efficiency coalitions by removing non-contributing agents. With tuned parameters, C-QLRG produces a near-winning coalition of size 11 and 99% reach on the synthetic network, surpassing the greedy baseline’s efficiency (size 12) despite a one-node coverage gap, while identifying the optimal winning coalition of size 1 on the Karate Club dataset, matching all baselines. The framework thus offers a principled trade-off between model fidelity and scalability, with the reward design choice being critical for practical deployment.
]]>Authors: Ahmed Ibrahim Ali AlSanousi Ahmed Serag
Agentic artificial intelligence (AI) offers new opportunities for intelligent clinical decision support, but deployment in emergency and crisis settings remains challenging because time-critical recommendations must often be generated under incomplete patient information and system constraints. Conventional clinical decision support systems rely on rule-based workflows that degrade when structured data are absent, while standalone language models lack coordination mechanisms to enforce mandatory safety checks. We present TriAgent, a multi-agent framework that unifies adaptive orchestration, iterative retrieval, embedded safety verification, and end-to-end auditability within a single crisis clinical decision support workflow. An Orchestrator Agent dynamically selects specialist modules for clinical assessment, retrieval, treatment planning, safety verification, and system coordination, with routing determined by model reasoning rather than fixed execution paths. A retrieval sub-agent performs iterative query refinement and relevance grading over 49,000 MIMIC-IV discharge notes, while medication-conflict screening and allergy-risk assessment are invoked in parallel only when clinically indicated. A Critique Agent reviews the full reasoning trace before recommendation finalization. In a retrospective evaluation on 1000 real emergency presentations under synthesized incomplete-information inputs, TriAgent achieved 85.0% critical-case recall and 65.7% overall triage accuracy, versus at most 14.7% and 43.4% for matched single-model and retrieval-only baselines, with safety checks executed on every continuation pathway and adaptive routing invoking only the modules each case required. These results support multi-agent orchestration as a promising design pattern for transparent and auditable AI in healthcare. These gains are internal system properties; clinical-safety benefit remains to be established through prospective, clinician-involved validation.
]]>Authors: Zewei Mi Minming Gu
Accurate detection of tea buds is a key technology for enabling automated tea harvesting. However, in natural environments, tea buds present challenges such as scale variation, dense distribution, and high similarity to the background, making it difficult for traditional methods to balance accuracy and efficiency. To address these issues, this paper proposes a lightweight detection framework, PCM-YOLO. The model introduces a cascaded gated feature modulation network into the YOLOv11 architecture, combining feedforward structures and gating mechanisms to selectively emphasize informative features, thereby improving tea bud detection performance. In addition, a feature-enhanced downsampling module is proposed, which employs a stepwise pooling-based feature enhancement mechanism to progressively expand the receptive field while preserving feature resolution, effectively incorporating multi-scale contextual information. Finally, a multi-scale feature enhancement module is designed to reduce the computational complexity of the model while maintaining detection performance as much as possible. Experimental results on public datasets demonstrate notable performance improvements over YOLOv11-N: Precision increases from 86.7% to 90.6% (an absolute increase of 3.9 percentage points), mAP50-95 increases by 1.6%, and the number of parameters is reduced by 20.6%. These results indicate that PCM-YOLO achieves a substantial reduction in model complexity while effectively improving detection accuracy, providing a feasible technical solution for deploying high-precision, real-time tea bud detection systems at the edge in tea plantation environments.
]]>Authors: Weifeng Zhai Jiahui Qiu Qingkuo Wang Binbin Li He Zhang
Industrial robot after-sales question answering often depends on multimodal evidence, such as error screenshots, interface displays, and wiring diagrams, which are difficult for conventional text-based retrieval-augmented generation (RAG) systems to exploit effectively. To address this issue, we design a dual-channel multimodal RAG system that converts image content into retrievable textual knowledge through the collaboration of optical character recognition (OCR) and structured semantic description. In the proposed system, OCR is used to extract explicit textual cues, such as error codes, parameter fields, and interface prompts, while expert-authored semantic descriptions complement implicit visual evidence, including device parts, fault phenomena, and contextual scene information. The transformed knowledge is further integrated into a hybrid retrieval pipeline that combines dense retrieval and BM25, followed by Reciprocal Rank Fusion (RRF) and Maximal Marginal Relevance (MMR) reordering to improve both relevance and contextual diversity. Experiments on a real-world industrial robot after-sales dataset show that the proposed method achieves an overall question-answering accuracy of 87.9%, outperforming the LLM-only baseline by 35.6 percentage points. For image-related questions, accuracy improves from 46.7% to 83.3%. These results indicate that the proposed framework provides a deployment-friendly and interpretable system-level alternative to end-to-end multimodal model fine-tuning for industrial after-sales question answering.
]]>Authors: Fnu Neha Deepshikha Bhati Deepak Kumar Shukla
Whole-slide imaging has transformed histopathology into a data-rich domain; however, many computational pathology models encode tissue morphology within latent representations, limiting interpretability, reproducibility, and generalization. This review positions histomics as an intermediate phenotype representation layer linking histological images with downstream clinical inference through structured descriptors of tissue morphology, spatial organization, and tissue architecture. Unlike prior reviews focused primarily on feature extraction or predictive performance, the study adopts a representation-centric perspective of histomics. A taxonomy of histomic features across biological scales is presented, and artificial intelligence frameworks, including machine learning, deep learning, weakly supervised learning, and multimodal approaches, are systematically examined. Key challenges, including segmentation dependence, feature instability, aggregation variability, and domain shift, are critically analyzed alongside emerging developments in foundation models, representation learning, and multimodal pathology. The review provides a unified framework for understanding histomic representations and identifies future directions for developing robust, interpretable, and generalizable computational pathology systems.
]]>Authors: Yi Zhang Shaoqi Dai Cheng Wang Xiuhe Li Jinhe Ran Guoqiang Zhu Wenbo Liu Shuyun Shi
The introduction of noise during image acquisition and transmission is inevitable, leading to a significant reduction in the accuracy of image processing tasks, such as target classification, localization, and recognition. To address this issue, this paper proposes a novel robustness-oriented pooling module called Gaussian adaptive pooling. Drawing on the principles of Gaussian filters, the method introduces a Gaussian weight for feature values in the pooling operation, thus integrating filtering and pooling in a novel manner. This approach is both lightweight and versatile, requiring no additional learnable parameters, and enables seamless integration into neural network architectures with pooling layers. Rigorous mathematical derivations and simulation experiments show that our proposed Gaussian adaptive pooling method surpasses conventional methods (average-pooling and max-pooling) in noise handling. Furthermore, its robustness is comparable to traditional pooling methods in addressing challenges such as rotations, scalings, and translations. Extensive evaluations across multiple computer vision tasks—including image classification (CIFAR-10/100), object detection (MS COCO and RTTS), and semantic segmentation (CamVid)—confirm its effectiveness. Specifically, under varying levels of noise and degraded conditions, Gaussian adaptive pooling achieves significant improvements in standard performance metrics compared to conventional pooling methods. For instance, it delivers notable quantitative gains across different tasks including up to a 12.67% increase in mean intersection over union on the CamVid dataset for semantic segmentation and a 1.1% mAP50 enhancement on the real-world RTTS dataset for object detection.
]]>Authors: George-Andrei Dima Ilie Cosmin Bilțan Mirabela-Melinda Medvei Luciana Morogan
Relation extraction is an important task for structuring information from unstructured text. However, the Romanian language still lacks dedicated datasets and benchmarks for this task. To address this gap, we introduce RoRED, a Romanian relation extraction dataset built by combining two complementary data construction strategies: translating existing high-quality English resources and applying distant supervision to native Romanian Wikipedia data. We leverage a powerful open-source large language model to automatically translate English examples into Romanian. For the native subset, we align Romanian Wikipedia entities with Wikidata relations to obtain naturally occurring Romanian examples. To better reflect real-world relation extraction scenarios, we also introduce synthetic negative examples generated using existing Romanian named entity recognition models. Finally, we validate the dataset by fine-tuning and evaluating multiple baseline models. Our strongest model, LUKE-RoRED, achieves a macro-F1 score of 0.8744 on the RoRED test set, demonstrating that the dataset can support relation extraction for Romanian. Overall, RoRED provides a strong first native benchmark for Romanian relation extraction.
]]>Authors: Hendrik Möller Alina Dima Benjamin Keinert-Weth Robert Graf Matan Atad Johannes Paetzold Friederike Jungmann Rickmer Braren Florian Kofler Bjoern Menze Daniel Rueckert Jan S. Kirschke Hanna Schön
Thoracolumbar stump ribs are one of the essential indicators of thoracolumbar transitional vertebrae or enumeration anomalies. While some studies manually assess these anomalies and describe the ribs qualitatively, this study aims to automate thoracolumbar stump rib detection and analyze their morphology quantitatively. To this end, we train a high-resolution deep learning model for rib segmentation using nnUNet and achieve significant improvements over existing models (Dice score 0.997 vs. 0.779, p-value < 0.01). In addition, we employ a novel iterative algorithm and piecewise linear interpolation to estimate rib length, achieving a success rate of 98.2%. When analyzing morphological features, we show that stump ribs articulate more posteriorly at the vertebrae (−19.2±3.8 vs. −13.8±2.5 mm, p-value < 0.01), are thinner (260.6±103.4 vs. 563.6±127.1mm2, p-value < 0.01), and are oriented more downwards and sideways within the first centimeters in contrast to full-length ribs. We show that with partially visible ribs, these features can achieve an F1-score of 0.84 and an AUC of 0.98 in differentiating stump ribs from regular ones. We publish the model weights and masks for public use.
]]>Authors: Ibrahim Al Janabi Tyler Bland
Generative artificial intelligence (genAI) enables educators to build custom learning tools, but the feasibility and impact of educator-driven, AI-assisted development (“vibe coding”) in medical education remain unclear. This study describes the rapid development of a custom ECG learning application using Gemini 3.1 Pro, evaluates its association with exam performance using difference-in-differences (DiD) and triple-difference (DDD) analyses, and assesses student perceptions with the user version of the Mobile App Rating Scale (uMARS). The app was implemented at one WWAMI site (intervention) with five sites as controls; aggregate performance from two first-year medical student cohorts (E24 vs. E25) was analyzed, comparing ECG-focused (focal) to non-ECG (baseline) exam items. DDD effects were inconsistent across exams, with no overall pooled effect on focal performance relative to baseline versus controls. In contrast, students rated the app highly (overall uMARS 4.57/5), particularly for quiz customization and waveform annotations. These findings support the feasibility of rapidly building and deploying tailored educational tools via genAI-assisted workflows and suggest strong perceived usability and acceptability among students. However, the study did not demonstrate a definitive short-term learning effectiveness effect on exam performance. Vibe coding is therefore positioned as a practical model for faculty-driven, context-specific educational innovation that requires further evaluation across broader implementations.
]]>Authors: Sana Amreen Ahmed Khairy Fakeha Masood Ngan Chu Anju Paudel Abdelrahman Aly Mohamed Ayantoyinbo Oluwabusayomi Yossef Alnasser
Background: Radiology underpins diagnosis and treatment across pediatrics, yet most artificial intelligence (AI) tools are developed for adults and validated on adult datasets only. Of more than 200 AI systems cleared by the United States (U.S.) Food and Drug Administration (FDA), only about 3% include pediatric validation. Because children differ from adults in anatomy, physiology, pathology, epidemiology, and imaging protocols, adult-trained models often perform sub-optimally in pediatric settings. Methods: A narrative review of peer-reviewed literature from 2000 to 2025 was conducted using PubMed, MEDLINE, Google Scholar, and Scopus. Studies involving AI applications in pediatric X-ray, ultrasound, computed tomography (CT), magnetic resonance imaging (MRI), echocardiography, and point-of-care ultrasound with quantitative performance metrics were included. Findings were synthesized by imaging modality, clinical task, and differences between high-income countries (HICs) and low- and middle-income countries (LMICs). Results: AI demonstrated strong performance across multiple pediatric imaging tasks. In X-ray interpretation, AI detected fractures with area under the curve (AUC) values up to 0.96 (sensitivity, 90.8%; specificity, 88.7%). Pneumonia classification achieved 76.5% accuracy, and foreign body aspiration detection showed 95.3% specificity in HICs. In ultrasound, AI improved junior sonographers’ detection of intussusception (AUC 0.857 to 0.966) and reduced scan time by more than 50%. AI-assisted bone age estimation achieved a mean error of 0.39 years. In echocardiography, AI-derived ejection fraction showed excellent agreement with experts’ interclass correlation coefficient (ICC 0.983), and AI support improved atrioventricular septal defect detection (84.4% to 86.5%). In MRI, the use of AI enhanced lesion detection and supported quantitative analysis. Deep-learning models trained on routine T1- and T2-weighted sequences predicted liver stiffness across multi-site datasets, while advanced neuroimaging pipelines improved the identification of subtle epileptogenic lesions that are often missed on conventional pediatric MRI. However, adult-trained models showed limited generalizability to children. Still, excluding children under the age of two years improved the reading accuracy of pediatric chest X-rays (CXRs) by adult-trained models from 88% to 97%. AI faces challenges beyond the development of age-specific models. Substantial heterogeneity, limited pediatric-specific datasets, and unresolved medicolegal responsibility further restrict adoption worldwide. Challenges are amplified in LMICs, where unstable electricity, limited radiology resources, weak digital infrastructure, and scarce pediatric providers limit implementation. Additionally, many large language models underperform and lack inclusive algorithms suitable for pediatric radiology in many LMICs. Conclusions: AI can enhance diagnostic accuracy, efficiency, and access to pediatric imaging, particularly in resource-limited settings, through task-shifting and decision support. However, it cannot replace pediatric radiologists as of today. Safe adoption requires pediatric-specific model development, standardized validation metrics, diverse datasets that include LMIC populations, stronger digital infrastructure, robust radiologist training in AI capabilities, and the establishment of clear guidelines and medicolegal policies.
]]>Authors: Veronica Manescu (Paltanea) Iosif-Vasile Nemoianu Gheorghe Paltanea Iulian Antoniac Aurora Antoniac Alexandru Streza Gabriel Cristescu Costel Paun Adrian-Vasile Dumitru
Background: Biodegradable implants manufactured from Mg-based alloys are one of the most commonly used in orthopedics. However, their overall clinical acceptance is influenced by their fast corrosion speed and hydrogen emission. Based on an innovative manufacturing route previously described, this study introduces a preliminary proof-of-concept for a Gemini-assisted Digital Twin (Gemini-DT),which is an AI-augmented in silico framework designed to consider a MgF2 conversion coating on the implant surface and to model the synchronization of the degradation process with new bone formation. Methods: Based on the integration of experimental data for Mg-Nd and Mg-Zn alloys and by considering the implant geometry and coating formation, we developed, in collaborative work with LLM Gemini 1.5 Flash (Google), a four-module cognitive framework (surface thermodynamic synergy (Module 1), degradation analysis and alloy extract concentration management (Module 2), micro-channel fluidics and mechanical stability (Module 3), and bio-mechanical synchronization and regenerative evaluation (Module 4)) to evaluate simulated implant behaviors). Results: Using a 10,000 iteration Monte Carlo stability simulation, the model demonstrated a potential 12% reduction in false-negative design screening errors compared to rigid rule-based systems, achieving strong internal decision consistency in sustaining the mandated parametric compliance window. Computational verification supports the projected biocompatibility trends of Mg-Zn alloys, as previously demonstrated in our in vivo studies. Conclusions: Our research leads to a consistent computational architecture dedicated to Mg-based implants and offers a robust platform for virtual design and optimization. These observations suggest that the developed model can recover viable designs, whereas traditional linear models may reject them.
]]>Authors: Miguel G. Juarez Jaime Cerda Alejandro Zamora-Mendez Jose Ortiz-Bejar Juan Carlos Silva-Chavez
Modern power systems are becoming increasingly complex due to the rapid integration of renewable energy sources, the widespread use of nonlinear power-electronic devices, and the deployment of microgrids operating in parallel with conventional power grids. These evolving conditions intensify the occurrence of diverse and highly complex power quality disturbances (PQDs), demanding accurate and computationally efficient monitoring strategies. This paper presents a novel multi-stage hierarchical framework for PQD detection and classification, comprising an initial training stage with a dedicated 1D Convolutional Neural Network (1D-CNN), a transfer learning stage, and a subsequent fine-tuning stage. The proposed approach operates directly on raw voltage waveforms, eliminating the need for any signal preprocessing, as the CNN performs internal feature extraction. The framework is evaluated using a comprehensive dataset that includes synthetic signals, Matlab/Simulink (version R2022a) time-domain simulations, and real voltage sag events. Additionally, up to 29 types of disturbances, including complex multi-event combinations defined by the IEEE-1159 Standard, are generated using the PQ-SyDa toolbox. The proposed model achieves an F1-score of 97.8% using a three-cycle analysis window and further improves to 98.86% when five cycles are used. These results highlight the robustness and generalization capability of the proposed approach for the real-time PQD monitoring task in modern electrical networks.
]]>Authors: Priyadarshini Raghavendra Manob Jyoti Saikia
This review examines the evolution and architectural foundations of agentic artificial intelligence (AI), with a focus on collaborative multi-agent systems for complex task execution. The paper analyzes the core components, agent architectures, coordination mechanisms, application domains, and deployment challenges that enable autonomous reasoning and decision-making in real-world environments. To complement the survey, a comparative cryptocurrency market analysis case study is conducted using CrewAI, LangChain, and LangGraph focusing on workflow orchestration characteristics such as tool invocation, task transitions, orchestration depth, and memory integration. The findings are further supported by evidence from real-world financial applications reported in the literature, indicating productivity gains of 50–80% in financial data tasks and up to 20% improvement in stock prediction accuracy, highlighting the growing impact of multi-agent AI systems in market intelligence. The study highlights how architectural design choices influence reasoning continuity, coordination behavior, scalability, and system reliability, providing practical guidance for the design and deployment of agentic AI systems in complex, data-intensive domains.
]]>Authors: Yusaku Fujii
Objectives: This paper presents a principle and framework for generating social trust in AI outputs as an institutional structure rather than an ethical declaration. Sound technical design alone does not guarantee the institutional trust required to establish social measurement. What is needed is not a declaration of trust but the construction of an infrastructure that supports it. Methods: First, the Extended Kelvin Principle is derived by prepending to Kelvin’s measurement–understanding–control chain the links “no social trust without trust infrastructure; no legitimate social measurement without social trust.” Infrastructure-scale trust requires not declarations but verifiability, recordability, and auditability. Just as GUM and calibration infrastructure underpin trust in measured values, AI output governance requires GLO, a common language for expressing output legitimacy, implemented by a VRAIO-type infrastructure. GLO treats an output candidate as a “claim” and declares the rule-conformity of its purpose and content as a legitimacy confidence L, derived from a fact-based argument accompanied by a legitimacy budget. Results: VRAIO integrates declaration, rule verification, tamper-resistant recording, and independent auditing. A sealed, deterministic verifier makes L reproducible: computational falsity is caught by re-computation, factual falsity by checking authoritative records, and severe sanctions render false declaration irrational. Conclusions: GLO is not a mere AI version of GUM but a common language for an underdeveloped domain, whose effectiveness depends on connection to an enforceable output-governance infrastructure.
]]>Authors: Feng Guo Xunhuang Wang Fumin Zou Lei Zou Tao Fang Xueming Wu Haocai Jiang Jianqing Weng
Accurate traffic flow prediction is fundamental to Intelligent Transportation Systems (ITSs), critical for transportation management and logistics. Despite advances in spatio-temporal prediction methods, existing approaches suffer from two key limitations: (i) multi-scale fusion methods inadequately capture hierarchical constraints between cross-scale features, and (ii) models rely on single spatio-temporal views, neglecting multi-source relationship complementarity. To address these issues, we propose ST-MAFNet, a spatio-temporal multi-scale adaptive fusion network comprising three key components, specifically, a Cross-Scale Hierarchical Anchoring strategy (CSHA) that anchors short-term predictions with multi-scale temporal patterns to mitigate noise; a Dual Spatial Perception Module (DSPM) that learns node heterogeneity and dynamic correlations through node embeddings and adaptive graph attention; and a Spatio-Temporal Adaptive Fusion Module (STAFM) that captures time-varying connectivity by integrating multi-scale temporal features with multi-source spatial relationships. Experiments on four real-world datasets demonstrate that ST-MAFNet is particularly effective for short-term traffic forecasting. Compared with the best previously reported MAE results, ST-MAFNet reduces MAE by 2.95%, 1.43%, 1.25%, and 0.37% on PEMS03, PEMS04, PEMS07, and PEMS08, respectively, and achieves the best or second-best performance on most evaluation metrics.
]]>Authors: Ainhoa Serna Jon Kepa Gerrikagoitia Juan de Oña
This study evaluates zero-shot domain transfer for multilingual sentiment analysis in sustainable urban mobility using XLM-RoBERTa, a transformer pre-trained on social media data and applied to transport reviews without task- or domain-specific fine-tuning. Starting from a manually annotated English corpus of 375 transport-related user reviews, we created sentence-aligned translations in Spanish, French, German, and Italian, yielding a multilingual evaluation dataset of 1875 instances. Results show that the model assigns consistently high confidence to polarized content (mean: 0.76–0.85) and lower confidence to neutral or ambiguous expressions (0.58–0.65), with visible but preliminary cross-lingual variations that require further linguistic validation. Confidence scores are treated as diagnostic indicators of model certainty, not as evidence of correctness or calibration. A qualitative analysis of 113 categorized low-confidence predictions identifies six recurring linguistic patterns associated with model uncertainty (led by translation drift, mixed sentiment, and idiomatic expressions) with substantial inter-annotator agreement (κ = 0.664). By releasing the annotated multilingual dataset and code publicly, this work provides a reproducible exploratory evaluation framework for annotation-scarce, domain-specific multilingual NLP.
]]>Authors: Tea Krčmar Dina Šabanović Mirko Köhler Ivica Lukić
Heterogeneous anomaly detection ensembles improve robustness by combining complementary detectors, yet existing approaches often rely on heuristic detector selection, fixed contamination assumptions, and equal weighting. We investigate whether compact ensembles of complementary detectors can outperform substantially larger heterogeneous configurations through diversity-aware weighting and adaptive contamination estimation. Experiments on 22 benchmark datasets show that a compact ensemble of four complementary classical detectors outperforms an eleven-detector ensemble containing deep learning components, while requiring only 13.8% of the computational cost. Across the benchmark, the proposed ensemble variants achieve strong rankings while remaining competitive with the strongest individual detectors (Friedman χ2=71.58, p<0.001). These findings suggest that detector diversity, rather than ensemble size or architectural complexity, is the primary driver of robust unsupervised anomaly detection performance in resource-constrained environments.
]]>Authors: Shunping Niu Kuo Chi Ting Su Yongqin Yang Jiabao Gao
Conventional recommender systems often rely on shallow collaborative signals, which limits their performance under sparse and popularity-skewed conditions. To address this, we propose a knowledge-aware framework that combines an item hypergraph induced by user interaction histories, a top-k user similarity graph, and one-hop, relation-aware knowledge-graph aggregation. The hypergraph branch learns high-order item co-occurrence representations, which are aggregated into initial user vectors and then refined through user similarity propagation. On the item side, user-conditioned relation attention aggregates one-hop KG neighbors to produce semantic item representations. User and item representations are fused by an MLP scorer, and a lightweight popularity-aware post-scoring adjustment can optionally be applied to moderate head-item dominance. Experiments on MovieLens-1M, Last.FM and Book-Crossing show strong performance among the compared baselines in AUC, ACC, and Recall@K.
]]>Authors: Alberto Gudiño-Ochoa Eduardo Ruiz-Velázquez Julio Alberto García-Rodríguez Raquel Ochoa-Ornelas Sofia Uribe-Toscano
Non-invasive blood glucose estimation from exhaled breath has been proposed as a painless alternative to repeated capillary measurements; however, performance evaluation remains challenging in small-sample settings. This study investigates the estimation of blood glucose from human breath using volatile organic compound (VOC) signals acquired with an electronic nose. Responses from three metal-oxide sensor channels sensitive to CO, alcohol, and acetone were collected from 58 individuals, with one measurement per subject, and analyzed using strictly patient-level five-fold cross-validation, in which test folds comprised only real subjects. Two experimental factors were examined. First, model performance was evaluated with and without an additional interpretable alcohol–acetone log-ratio capturing relative variation between compounds. Second, model training was performed using either real data only or fold-wise tabular synthetic augmentation generated via a Gaussian copula fitted exclusively on training subjects, while evaluation remained strictly real-only. Under real-only training, classical machine learning models achieved the lowest prediction errors (approximately 6–7 mg/dL), whereas under synthetic augmentation FTTransformer was the best-performing deep learning model. This findings should be understood as a constrained proof-of-concept analysis rather than as evidence of diagnostic capability or clinical readiness.
]]>Authors: Panagiotis Douros Konstantinos Kasaras Konstantinos Milioris
Background: The rapid advancement of generative artificial intelligence is transforming music composition from an exclusively human-centric activity into a hybrid human–algorithmic domain. Despite technological progress and growing commercial integration, consumer acceptance of AI-generated music remains empirically underexplored. Methods: This study formulates and empirically evaluates a multidimensional theoretical model integrating nine frameworks—including UTAUT2, parasocial interaction theory, anthropomorphism theory, authenticity theory, and innovation resistance theory—through a quantitative cross-sectional survey of 466 young adults aged 17–28. Confirmatory factor analysis and multiple regression analysis (with robust standard errors) were employed. Results: The model explained 63.6% of the variance in behavioral intention (R2 = 0.636). Five constructs emerged as significant predictors: hedonic motivation (β = 0.136, p = 0.017), parasocial relationships (β = 0.121, p = 0.002), social influence (β = 0.126, p = 0.002), performance expectancy (β = 0.102, p = 0.019), and innovation resistance (β = −0.089, p = 0.029). Authenticity concerns, ethical AI concerns, anthropomorphic perceptions, and technological substitution fears were non-significant in the multivariate model. Conclusions: Young consumers’ acceptance of AI-generated music is primarily driven by experiential, social, and relational factors rather than ethico-cultural concerns. These findings have substantive implications for creative industries navigating algorithmic cultural production.
]]>Authors: Gulnaz Zakariya Aiman Moldagulova Nor’ashikin Ali
Point-of-sale (POS) consumer credit represents the most rapidly expanding retail-lending channel within the emerging Eurasian markets, necessitating a stringent operational framework for the underwriting model: the decision must be rendered within a mere few hundred milliseconds during the in-store checkout process, while the inputs are constrained to what the application XML is capable of conveying. This research endeavors to develop, internally validate, and operationally delineate a hybrid, explainable artificial intelligence framework aimed at POS credit scoring within the production portfolio of Kazakhstan’s largest second-tier bank. The architectural framework is delineated along two orthogonal dimensions—client tenure and decision-making channel—resulting in the formulation of three distinct production models: two transparent Weight of Evidence–Logistic Regression scorecards tailored for the real-time channel, and one isotonically-calibrated stacked ensemble (comprising LightGBM, CatBoost, and a three-layer neural network) designated for the batch channel. The selection of hyperparameters was conducted utilising Bayesian optimization within the context of stratified five-fold cross-validation. The digital scorecards achieve an area under the receiver operating characteristic curve (AUROC) of 0.847 and 0.835, whereas the offline ensemble enhances performance to an AUROC of 0.918, accompanied by a Kolmogorov–Smirnov statistic of 0.682 and a Gini coefficient of 0.836. The population stability indices persist below the threshold of 0.07, while isotonic recalibration effectively reduces the Brier score by 18%. Furthermore, an extensive examination of fairness demonstrates variations in approval rates within a margin of ±1.2 percentage points—and equalised-odds gaps below 1.5 percentage points in the true-positive rate and 0.7 percentage points in the false-positive rate—across multiple demographic factors such as gender, age, and distinctions between urban and rural classifications, thus establishing an artificial intelligence framework that is both regulatorily compliant and interpretable, aligning with the directives set forth by the Agency of the Republic of Kazakhstan for Regulation and Development of the Financial Market.
]]>