<?xml version="1.0" encoding="utf-8"?>
    <rss version="2.0">
      <channel xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <title>Frontiers in Artificial Intelligence | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence</link>
        <description>RSS Feed for Frontiers in Artificial Intelligence | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-06-03T12:13:17.647+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1818128</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1818128</link>
        <title><![CDATA[Evidence-based AI: from trailblazer to trustblazer?]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Perspective</category>
        <author>Thomas Luechtefeld</author><author>Thomas Hartung</author>
        <description><![CDATA[Agentic AI systems can plan, call tools, and coordinate specialized sub-agents, enabling multi-step scientific workflows that exceed what single-model text generation can reliably deliver. Yet in high-stakes domains such as regulatory science and toxicology, fluent outputs are not sufficient: adoption hinges on traceability, reproducibility, context-of-use validity, and explicit uncertainty communication. This perspective argues that evidence-based medicine and evidence-based toxicology provide a mature epistemic scaffold for making agentic AI trustworthy by design. We propose an Evidence-based Agent Stack that decomposes end-to-end tasks into protocolized roles (question framing, retrieval, screening, extraction, risk-of-bias appraisal, synthesis, mechanistic/causal integration, uncertainty assessment, and evidence-to-decision translation) with mandatory provenance and versioning. Anchoring agentic workflows in systematic review practice, risk-of-bias frameworks, and emerging regulatory principles (e.g., TREAT and e-validation) can turn “trailblazing” AI into “trustblazing” AI: systems whose outputs are auditable, updateable, and aligned with decision accountability.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1796682</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1796682</link>
        <title><![CDATA[Sovereign AI supercomputers: a global landscape review of unprecedented biomedical research infrastructure]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Review</category>
        <author>Lansaol Yang</author><author>Michael E. Bryan</author><author>Eduardo Veiga</author><author>Ian Lowenhoff</author><author>Alex Wan</author><author>Isam Mina</author><author>Tracey Allen</author><author>Javier Antonio Alfaro</author><author>Gareth Bloomfield</author><author>Julian Beach</author><author>Kristen Dahlgren</author><author>Nick K. Davis</author><author>Elisa Fontana</author><author>Spyridon Gennatas</author><author>Qamar Ghafoor</author><author>Franck Housseau</author><author>Daniel Lubelski</author><author>Zhehao Zhang</author><author>Matt Hancock</author><author>William Ince</author><author>Dominic James</author><author>Sam Khan</author><author>Victoria Kunene</author><author>John McGrane</author><author>Gerard Cathal Millen</author><author>Benjamin Moxley-Wyles</author><author>David Narganes-Carlon</author><author>Miranda Payne</author><author>Paul J. Ross</author><author>Rene Roux</author><author>Michael Rowe</author><author>Rebecca Lee</author><author>Jerry S. H. Lee</author><author>Justin K. H. Liu</author><author>Deepak Aggarwal</author><author>Aaron B. S. Teoh</author><author>Chrissie Thirlwell</author><author>Michael Tilby</author><author>Stefan Symeonides</author><author>Isabella Watts</author><author>David B. Agus</author><author>Santa J. Ono</author><author>Tim Elliott</author><author>Paul Calleja</author><author>Lennard Y. W. Lee</author>
        <description><![CDATA[Artificial intelligence (AI) has rapidly become the focal point of global governmental attention and investment. Nations are launching AI for science strategies on a scale comparable to historic endeavors such as Apollo and the Manhattan Project. These coordinated programs carry profound promise for people living with cancer, for those at risk of disease and for transformative public benefit. Central to this transformation is the rise of sovereign AI supercomputers which are fundamentally reshaping biomedical research. These publicly owned systems provide secure, large-scale computational capacity, enabling integration of complex health data and rapid analysis that was previously constrained. This review examines the geographic distribution, technical capabilities, and biomedical applications of these infrastructures. Key computational workloads that now benefit significantly from AI implementations include cancer imaging and diagnosis, personalized treatments, whole-genome and single-cell level analysis, and computational drug discovery. This approach has supercharged our efforts at the United Kingdom’s Cancer Vaccine AI & Supercomputing Project, our flagship national initiative to create new AI foundation models to accelerate the development of tools to establish immunity from cancer. In addition, this review evaluates governance models that safeguard patient privacy and intellectual property as well as measures that promote international collaboration while preserving compliance with regional regulations and make safer, more precise and effective treatments for public benefit. Substantial challenges exist, however, including inequitable resource availability, heterogeneous data standards and regulatory frameworks, and unbalanced computational expertise impeding the effective use of sovereign compute. Global collaborations are key to providing equitable access to advanced analytics, shortening the path from bench to bedside, and developing critical innovative tools for people affected by cancer.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1804943</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1804943</link>
        <title><![CDATA[CITADEL: a post-quantum secure blockchain framework for privacy-preserving electronic health records with temporally-partitioned federated learning]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Nagaraj Segar</author><author>Vijayarajan Vijayan</author>
        <description><![CDATA[IntroductionElectronic health records (EHRs) increasingly anchor clinical decision support and population-scale analytics, yet their concentration of sensitive information amplifies disclosure risk, widens the attack surface, and faces emerging threats from quantum computing. Existing frameworks fail to simultaneously address privacy preservation, quantum-resistant security, and cross-institutional federated learning.MethodsWe introduce CITADEL (Cryptographically Integrated Temporal Architecture for Distributed EHR Ledger), integrating five co-designed components: NIST-standardized CRYSTALS-Kyber (ML-KEM-768) and CRYSTALS-Dilithium (ML-DSA-65) post-quantum cryptography via the validated pqcrypto library; a genomic-aware privacy engine with beacon query protection and calibrated randomized response; temporally-partitioned federated learning with hospital-specific weighted aggregation; multi-modal health data tokenization; and an adaptive regulatory compliance engine for HIPAA and GDPR. Evaluation used a synthetic EHR dataset comprising 5,000 patients across 10 healthcare institutions, with 30-day hospital readmission as the primary prediction task.ResultsCITADEL achieves 84.5% accuracy and 0.866 AUC-ROC, exceeding nine baselines including centralized neural networks and differentially-private federated learning. Privacy metrics include k-anonymity of 13, l-diversity of 2.0, 99.0% linkage attack resistance, 42.2% attribute inference resistance, and 100% correlation preservation. The ledger sustains 285.3 transactions per second with ML-DSA-65 signing in 2.16 ms and verification in 0.46 ms. Multi-seed evaluation confirms robustness (accuracy 0.854 ± 0.012, AUC-ROC 0.880 ± 0.014).DiscussionCITADEL demonstrates that privacy preservation, quantum-resistant security, and usable federated analytics can be reconciled within one cohesive architecture. Results suggest a practical route to healthcare data management that remains credible in a post-quantum computing era and compatible with decentralized governance.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1851069</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1851069</link>
        <title><![CDATA[A systematic pipeline for diagnosing and reducing gender-stereotype bias in Japanese PLMs for sentiment analysis]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yang Liu</author><author>Ziyang Li</author>
        <description><![CDATA[Pre-trained language models (PLMs) are widely used in sentiment analysis, but they may inherit gender-stereotypical bias from large-scale text corpora and transfer such bias to downstream sentiment predictions. Despite growing attention to gender-stereotypical bias in PLMs, existing studies predominantly focus on English corpora and static word embeddings, limiting understanding of how such bias affects sentiment analysis models and the effectiveness of mitigation strategies. In this study, we present a three-stage task-oriented pipeline for diagnosing, mitigating, and evaluating gender-stereotypical bias in Japanese PLMs for sentiment analysis tasks. Specifically, the proposed framework diagnoses bias based on pro-stereotypical (PS), anti-stereotypical (AS), and neutral test sets, which are constructed to compare stereotype-aligned, stereotype-violating, and gender-neutral contexts under the same sentiment analysis setting. We further introduce two complementary evaluation measures, the Stereotype Bias Index (SBI) and Gender Sentiment Bias (GSB), to quantify stereotype-level bias between PS and AS samples, as well as sentiment prediction differences among male, female, and neutral groups. To mitigate bias, the framework performs debiasing fine-tuning using gender-swapped training data and then quantitatively evaluates bias reduction while monitoring sentiment classification performance. Experimental results on three Japanese BERT-based sentiment analysis models demonstrate that the proposed pipeline substantially reduces gender-stereotypical bias. For the SBI metric, the bias magnitude is reduced by 97.4%, 70.0%, and 76.9% for Tohoku BERTBASE, Tohoku BERTBASE(chABSA), and Tohoku BERTBASE(JSPD), respectively. For the GSB metric, the bias magnitude is consistently reduced across gender-group comparisons, with reduction rates ranging from 33.3% to 98.8%. Meanwhile, sentiment classification performance is maintained or slightly improved after debiasing.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1765191</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1765191</link>
        <title><![CDATA[Collaborate and explain on-the-fly: knowledge-based reasoning and learning in ad hoc teamwork]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Hasra Dodampegama</author><author>Mohan Sridharan</author>
        <description><![CDATA[This paper focuses on ad hoc teamwork, the problem of enabling an AI agent to collaborate with other agents without prior coordination. Methods considered state of the art for ad hoc teamwork formulate it primarily as a learning problem, using a large labeled dataset of different situations to model the action choices of other agents (or agent types) and determine the actions of the ad hoc agent. Such datasets are not readily available in practical domains, and these methods lack transparency and make it difficult to rapidly revise existing knowledge (or models) in response to changes in the domain, team composition, or agents' capabilities. Our architecture for ad hoc teamwork embeds the principles of refinement, ecological rationality, interactive learning, and explainable agency, leveraging the complementary strengths of knowledge-based and data-driven methods for reasoning and learning. Specifically, for any given goal, our architecture enables an ad hoc AI agent to determine its actions through non-monotonic logical reasoning with: (a) prior domain-specific commonsense knowledge; (b) models learned and revised rapidly to predict the behavior of other agents; and (c) anticipated abstract future goals based on generic knowledge of similar situations in a pretrained Large Language Model. In addition, the ad hoc agent processes natural language descriptions and observations of other agents' behavior, using a combination of a pretrained Large Language Model and decision-tree induction to incrementally acquire and revise knowledge in the form of objects, actions, and axioms that govern domain dynamics. Furthermore, the ad hoc agent generates relational descriptions as on-demand explanations of its decisions and beliefs, and those of other agents, in response to various types of questions. We ground and experimentally evaluate the capabilities of our architecture in VirtualHome, a realistic, physics-based 3D simulation environment. We demonstrate reliable, efficient, transparent, and scalable performance, providing a substantial improvement in performance compared with a purely knowledge-based baseline, and comparable or better performance than a purely data-driven baseline while using orders of magnitude fewer resources.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1786679</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1786679</link>
        <title><![CDATA[Deep learning driven colorectal polyp analysis: a review of detection, classification and segmentation methods]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Systematic Review</category>
        <author>Divya S</author><author>Sudha M</author>
        <description><![CDATA[Colorectal polyps are key determinants of colorectal cancer. Their accurate detection during colonoscopy has been a technically challenging work due to differences in shape size imaging conditions and texture. Emerging advances in Artificial Intelligence predominantly in deep learning have been making significant changes in the automatic detection and classification of polyps. This review presents a systematic and in-depth analysis of artificial intelligence-based methods for colorectal polyp detection classification and segmentation. Publicly available datasets are extensively reviewed along with data pre-processing and augmentation techniques that highlights low contrast noise and class imbalance. The review also investigates about the present state-of-the-art models for all three tasks. It is based on architecture designs performance trends and relative strengths. A thorough assessment has been made for the standard performance metrics used in existing literature for fair and consistent benchmarking. Finally existing gaps and future research paths have been discussed with an objective to fill the performance-translation gaps between experimental performance and clinical deployment. This review gives a structured reference for AI-based colorectal polyp analysis.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1868452</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1868452</link>
        <title><![CDATA[Compressed professionalization in informal economies: a socio-technical analysis of youth-led artificial intelligence adoption in the Democratic Republic of the Congo]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Delphin B. Kyubwa</author>
        <description><![CDATA[Artificial intelligence (AI) is increasingly shaping development trajectories across the Global South, yet limited attention has been paid to how AI is appropriated within highly informal and institutionally fragile economies. This article advances a conceptually driven, theory-building analysis supported by qualitative field insights used as illustrative grounding rather than for statistical generalization. Drawing on 125 semi-structured interviews conducted in Kinshasa, Lubumbashi, and Goma, and integrating Information and Communication Technologies for Development (ICT4D), socio-technical systems theory, and the capability approach, the study examines how infrastructural constraints, fragmented governance, and uneven skill ecosystems interact with youth-driven innovation to shape AI adoption in the Democratic Republic of the Congo (DRC). The article introduces compressed professionalization, defined as the accelerated acquisition and immediate market enactment of professional-level digital capabilities outside formal institutional pathways. Empirical observations show that youth mobilize AI tools for translation, content creation, customer engagement, and micro-entrepreneurial activities, enabling partial and situational approximation of selected formal-sector practices. The analysis further conceptualizes AI as a conditional capability amplifier, expanding agency while producing uneven inclusion shaped by disparities in connectivity, skills, and infrastructure. Rather than following policy-led or infrastructure-first trajectories, AI adoption emerges through hybrid socio-technical interactions between bottom-up youth innovation and weakly coordinated institutional frameworks. The article concludes by proposing a Strategic Action Framework to support more inclusive and context-responsive AI ecosystems. While grounded in the DRC, the findings offer broader insights into AI adoption dynamics across informal economies in Sub-Saharan Africa and beyond.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1764233</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1764233</link>
        <title><![CDATA[Machine learning analysis of student consumer choices for pet supplements]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ta-Chen Chen</author><author>Hui Yu Chung</author><author>Fu Shih Chen</author>
        <description><![CDATA[BackgroundMotivated by the rising global trend of veterinary Complementary and Alternative Medicine (CAM) usage and a specific data gap in Taiwan, this study investigates the consumption behavior of future pet owners.MethodsA cross-sectional survey was conducted among Taiwanese medical university students using a validated online questionnaire. Beyond traditional descriptive statistics, this study employed machine learning techniques to analyze owner demographics, pet characteristics, and determinants of CAM usage.ResultsData analysis revealed a strong correlation between positive owner perceptions—specifically satisfaction, belief in benefits, and understanding—and targeted CAM application. A decision tree model successfully identified “overall satisfaction” as the primary splitting criterion for user segmentation, followed by belief and understanding. Predictive modeling demonstrated high accuracy in identifying usage motivations for joint and digestive health, though predicting “immune system boosting” proved more complex due to behavioral variability.ConclusionOwner satisfaction is the critical predictor of CAM usage patterns. While the predictive model for specific conditions, such as joint and digestive health, yielded high accuracy (AUC > 0.93), these findings should be interpreted as an exploratory framework given the pilot nature of the study and the limited sample size (n = 41). These findings suggest that veterinarians and industry stakeholders should adopt data-driven communication strategies focusing on transparency and satisfaction.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1825655</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1825655</link>
        <title><![CDATA[Attention integrated deep learning models for interpretable multi-class IoT intrusion detection using SHAP]]></title>
        <pubdate>2026-06-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ramakrishnan Raman</author><author>Rahul Kumar</author><author>Benson Edwin Raj</author><author>Vishwesh Akre</author>
        <description><![CDATA[The rapid growth of the Internet of Things (IoT) has increased the size, complexity, and vulnerability of network traffic, making intrusion detection a critical factor of modern cybersecurity. Traditional intrusion-detection systems (IDSs) analyze handcrafted features and rules to detect emerging attack patterns. To address these limitations, deep learning frameworks, such as attention-enhanced one-dimensional (1D) Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), are analyzed for accurate and efficient multi-class attack detection. The proposed attention-enhanced 1D convolutional layers extract discriminative spatial features while focusing on the most relevant patterns within the network flow. The LSTM and GRU architectures analyzed long-range temporal dependencies present in sequential traffic data, enabling robust identification of subtle anomalies. The model was evaluated on two datasets: Hacking and Countermeasure Research Lab (HCRL) and the Kitsune IoT datasets, which represent diverse real-world benign and malicious traffic conditions. Experimental results demonstrate that the attention-enhanced 1D CNN achieved the highest performance of 98 and 87% on the Kitsune and HCRL datasets, respectively. SHapley Additive exPlanations (SHAP) based interpretability analysis shows how individual features contribute to predictions, highlighting the most significant features driving intrusion-detection decisions. The results confirm that incorporating attention mechanisms significantly enhances the discriminative capability, enabling more reliable classification of complex IoT attack types. The proposed approach effectively addresses key challenges in IoT intrusion detection by combining spatial and temporal deep learning components for deployment in intelligent real-time IoT network security systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1861536</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1861536</link>
        <title><![CDATA[Hierarchical context enhancement for long-tail entity retrieval augmented generation]]></title>
        <pubdate>2026-06-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yixuan Peng</author><author>Kewu Pan</author>
        <description><![CDATA[IntroductionRetrieval-Augmented Generation (RAG) in Domain-specific Question Answering (DSQA) often faces significant performance degradation due to semantic drift. Our analysis reveals that the main cause is the absence of a dedicated mechanism for handling low-frequency terms.MethodsMotivated by this observation, we propose a hierarchical context enhancement retrieval augmented generation (HCE-RAG). Specifically, in the indexing stage, we anchor low-frequency entities offline through entity-sensitive contextual tagging. During query processing, we perform minimal yet entity-focused query clarification via constrained query reflection. Finally, in the retrieval stage, we employ hybrid retrieval with RRF to balance contextual signals and exact word matching, thereby enabling robust identification of low-frequency entities.ResultsExperiments on a dedicated domain specific QA benchmark show that our method achieves strong results, delivering a 29 percentage point gain in Recall@10 on low-frequency entities.DiscussionNotably, the proposed approach is plug-and-play and can be directly integrated with existing state-of-the-art algorithms to improve their response accuracy in long-tail entity question answering.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1816292</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1816292</link>
        <title><![CDATA[Self-calibrating neuromorphic system for adaptive environmental sensing]]></title>
        <pubdate>2026-06-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Anantharaman Prasad</author><author>S. Sofana Reka</author><author>Prakash Venugopal</author>
        <description><![CDATA[Precision agriculture demands accurate, real-time environmental monitoring, conventional soil moisture sensors face critical issues such as long-term drift, high energy consumption, and limited adaptability to dynamic environmental changes. These limitations often lead to suboptimal irrigation decisions, wasted resources, and unreliable data, especially in remote or resource-constrained farming regions where frequent manual recalibration is impractical or impossible. This work addresses these challenges by introducing a novel self-calibrating neuromorphic system for adaptive soil moisture sensing. The system leverages Spiking Neural Networks (SNN) deployed on a low-power STM32H563ZI microcontroller. Our proposed solution autonomously recalibrates sensors to mitigate drift, significantly reduces energy consumption through event-driven computation, and adapts seamlessly to changing environmental conditions. The SNN model achieved a Mean Absolute Error (MAE) of 0.4557 and a Root Mean Squared Error (RMSE) of 0.5850, reducing baseline drift from 5.3% to 1.6% over a two-month deployment outperforming models like Isolation Forests and Autoencoders in predictive accuracy. This work significantly contributes to the growing field of neuromorphic computing in IoT applications, offering a scalable, low-power solution for precision agriculture and broader environmental monitoring. The demonstrated effective deployment of SNN-based learning mechanisms on low-constrained microcontroller hardware opens new avenues for resilient, decentralized intelligence in smart homes, wearables, and autonomous infrastructure inspection.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1814012</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1814012</link>
        <title><![CDATA[An integrated evolution-aware meta-learning framework with adversarial morphological augmentation for zero-day threat detections]]></title>
        <pubdate>2026-06-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Kavitha Lanka</author><author>Kareemulla Shaik</author>
        <description><![CDATA[IntroductionMost modern threat detection frameworks rely on fixed class definitions and retrospective signatures derived from historical data, limiting their ability to adapt to evolving attack behaviors. However, contemporary threats are increasingly polymorphic, adaptive, and include a growing number of zero-day attacks, making traditional pattern-matching approaches insufficient. With expanding attack surfaces across cloud environments, IoT systems, and heterogeneous infrastructures, detection systems must move beyond static similarity matching and instead account for the evolutionary nature of threats. Despite advances, both deep learning-based and traditional anomaly detection approaches face critical limitations. Deep learning models often suffer from data scarcity, concept drift, poor uncertainty calibration, and limited causal interpretability, while conventional anomaly detection methods exhibit slow training, weak generalization, and reduced adaptability to novel attack patterns.MethodsTo address these challenges, we propose M3-GAZE (Meta-Morphological GAN-Augmented Zero-day Detection Engine), which formulates threat detection as an evolutionary inference problem. The framework consists of five analytically distinct yet interdependent components. First, the Latent Morphology Spectrum Extraction (LMSE) module learns a continuous latent representation capturing structural invariants of threats across morphological variations, enabling a more flexible representation than discrete labels. This latent space is utilized by the Adversarial Evolutionary GAN (AE-GAN) to generate evolution-consistent synthetic samples that reflect unseen attack variations while minimizing unrealistic artifacts. These synthetic samples enhance the training of the Meta-Adaptation Framework for Threat Intelligence (MAFTI), which learns adaptation strategies across multiple evolutionary tasks, enabling accurate zero-day detection under limited data conditions. The enriched latent space further supports Adversarial Uncertainty Calibration Layer (AUCL), which evaluates epistemic uncertainty by introducing controlled adversarial perturbations. This transforms uncertainty into an action able early warning signal. Finally, the Causal Evolutionary Threat Graph Synthesizer (CETGS) constructs a temporal causal graph that explains threat evolution, detection decisions, and propagation dynamics.Results and discussionOverall, M3-GAZE improves zero-day recall, enhances few-shot adaptability, calibrates uncertainty, reduces false negatives, and provides interpretable, temporally aware threat detection.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1838463</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1838463</link>
        <title><![CDATA[Integrating DeepL Write and Claude AI to enhance argumentative writing competence of engineering students in India]]></title>
        <pubdate>2026-06-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>T. J. Divya</author><author>C. Alamelu</author>
        <description><![CDATA[IntroductionArtificial intelligence (AI) tools are increasingly applied to support second language writing. However, empirical research examining their potential to enhance the argumentative writing skills of STEM engineering students, particularly in the English as a Second Language (ESL) context in India, remains limited and requires more exploration.MethodThis study employed a quasi-experimental mixed-method design to investigate the effectiveness of two AI tools, DeepL Write and Claude AI, in improving argumentative writing among 40 first-year engineering students at a private university in Chennai, India. The Toulmin model of argumentation-based rubric was used to analyse the argumentative essays written by participants before and after the intervention using the AI tools. In the quantitative phase, paired-sample t-test and one-way ANOVA were employed to analyse differences in the pre-test and post-test essay scores. These results were further explained through comparative analysis of the essay scripts (n = 40) and thematic coding of participants’ semi-structured interview data (n = 23) in the qualitative phase.ResultsThe quantitative results demonstrate significant improvement in the AW skills of the participants (t = −43.72; df = 39; p-value < 0.001; f = 1.534; p > 0.05). The findings also indicated uniform improvements across all Toulmin components. The comparative analysis of the essays evidences the development of claims, data, warrants, backing and rebuttals, as well as improvements in linguistic features such as grammar, sentence structure and vocabulary. The thematic coding of the interview data yielded six themes related to perceived writing development, meta-cognitive awareness and the challenges and limitations of the tools.DiscussionThe findings indicate the pedagogical potential of integrating DeepL Write and Claude AI to improve argumentative writing skills of Indian ESL Engineering students. The findings also highlight the effectiveness of AI tools in improving linguistic features in students’ writing. By fostering responsible and informed use of digital technology in education, this study aligns with the Sustainable Development Goal (SDG) 4 (Quality Education).]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1824067</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1824067</link>
        <title><![CDATA[A lightweight CNN–transformer hybrid architecture with channel attention for real-time hazardous acoustic event detection]]></title>
        <pubdate>2026-06-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Aigerim Altayeva</author><author>Nurzhan Omarov</author>
        <description><![CDATA[IntroductionHazardous acoustic event detection is critically important for intelligent surveillance, emergency response systems, and public safety monitoring applications. Accurate and real-time identification of dangerous sound events such as explosions, alarms, screaming, and weapon-related sounds can significantly improve situational awareness and accelerate emergency response in safety-critical environments.MethodsThis study proposes a lightweight deep learning architecture for hazardous sound classification based on convolutional feature extraction and channel attention mechanisms. The proposed framework utilizes log-mel spectrogram representations as input and incorporates a TinyCNN backbone enhanced with squeeze-and-excitation channel attention modules to improve discriminative spectral feature learning while preserving computational efficiency. A custom balanced dataset consisting of eight hazardous acoustic classes, including crying, dog barking, emergency alarm, explosion, fire, glass breaking, screaming, and weapon-related sounds, was constructed with one thousand audio samples per class. The model was evaluated using accuracy, precision, recall, and F1-score metrics.ResultsExperimental results demonstrate that the proposed architecture achieves strong multi-class classification performance while maintaining real-time inference capability suitable for edge deployment scenarios. Quantitative evaluations confirm the effectiveness of the lightweight framework for hazardous acoustic event detection. Additional ablation studies indicate that the integration of channel attention mechanisms and spectrogram-based augmentation strategies substantially improves model robustness, feature discrimination, and generalization performance.DiscussionThe obtained findings demonstrate that the proposed lightweight channel-attention-enhanced architecture provides an efficient and reliable solution for real-time hazardous sound detection in intelligent monitoring and public safety systems. The combination of computational efficiency and robust classification performance highlights the suitability of the proposed framework for deployment in resource-constrained and edge-based environments.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1804284</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1804284</link>
        <title><![CDATA[Can small language models handle context-summarized multi-turn customer-service QA? A synthetic data-driven comparative evaluation]]></title>
        <pubdate>2026-06-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Lakshan Cooray</author><author>Deshan Sumanathilaka</author><author>Pattigadapa Venkatesh Raju</author>
        <description><![CDATA[Customer-service question answering (QA) systems increasingly rely on conversational language understanding. While Large Language Models (LLMs) achieve strong performance, their high computational cost and deployment constraints limit practical use in resource-constrained environments. Small Language Models (SLMs) provide a more efficient alternative, yet their effectiveness for multi-turn customer-service QA remains underexplored, particularly in scenarios requiring dialogue continuity and contextual understanding. In this study, we evaluate whether instruction-tuned SLMs, fine-tuned using parameter-efficient finetuning, can effectively handle context-summarized multi-turn customer-service QA while preserving contextual consistency, response quality and task relevance under computational constraints. We further investigate instruction-tuned SLMs for context-summarized multi-turn customer-service QA using a history summarization strategy to preserve essential conversational state and introduce a conversation stage-based qualitative analysis to evaluate model behavior across different phases of customer-service interactions. The main contributions of this work include the application of parameter-efficient fine-tuning to adapt SLMs for context-summarized multi-turn customer-service QA, a synthetic data construction pipeline for generating a context-summarized multi-turn QA dataset, and a structured evaluation framework combining quantitative metrics with human and LLM-as-a-judge assessments for customer-service QA evaluation. Nine instruction-tuned SLMs are evaluated against three commercial LLMs using lexical and semantic similarity metrics alongside qualitative assessments, including human evaluation and LLM-as-a-judge methods. Results show notable variation across SLMs, with some models demonstrating near-LLM performance, while others struggle to maintain dialogue continuity and contextual alignment. These findings highlight both the potential and current limitations of low-parameter language models for real-world customer-service QA systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1834985</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1834985</link>
        <title><![CDATA[A multi-agent RAG system for generating SCORM courses from enterprise documents]]></title>
        <pubdate>2026-06-01T00:00:00Z</pubdate>
        <category>Technology and Code</category>
        <author>Gulshat Amirkhanova</author><author>Bauyrzhan Amirkhanov</author><author>Alikhan Amirkhanov</author><author>Ramilya Aubakirova</author>
        <description><![CDATA[Corporate onboarding requires the effective transfer of complex organizational knowledge embedded in internal policies and procedural documents; however, existing artificial intelligence (AI)-driven course generation systems primarily target academic or public knowledge domains. This gap limits the scalability and consistency of enterprise training, particularly in regulated environments where factual accuracy is critical. In this study, we present a multi-agent pipeline that automatically generates Sharable Content Object Reference Model (SCORM) 1.2-compliant e-learning courses from heterogeneous enterprise documents using large language models and retrieval-augmented generation (RAG). The system integrates four stages: semantic document ingestion with structure-aware chunking and embedding, an autonomous ReAct-based architect agent for course design, a parallel content generation pipeline combining multi-query retrieval and neural reranking, and standards-compliant SCORM packaging for deployment in learning management systems. Evaluated using real-world occupational safety documents, the system produced a complete multi-module course with structured lessons and assessments within minutes, demonstrating end-to-end automation of instructional design grounded exclusively in source materials. By ensuring traceability of generated content to organizational knowledge, the approach reduces the risk of hallucinations.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1796177</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1796177</link>
        <title><![CDATA[Dual-band stub-loaded monopole antenna with bandwidth enhancement using weighted figure-of-merit optimization]]></title>
        <pubdate>2026-06-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jun-Jiat Tiang</author>
        <description><![CDATA[This paper presents a compact dual-band stub-loaded T-monopole antenna optimized for WLAN (2.4 GHz) and 5–6 GHz (5G/IoT) applications using a weighted figure-of-merit (FOM) and artificial neural network (ANN) surrogate modeling. Low- and high-band stubs enable independent resonance control, achieving −10 dB impedance bandwidths of 1.7–2.7 GHz (45.5%) in the lower band and 5.1–5.9 GHz (14.5%) in the upper band, with return loss depths exceeding −20 dB at resonances. This outperforms a conventional reference design (22.2% lower/9.0% upper) and prior ML-optimized stub-loaded monopoles. The weighted FOM prioritizes upper-band performance for high-data-rate needs (weights w₂ = w₄ = 1.5). An ANN surrogate, trained on 210 HFSS-simulated samples, yields R2 > 0.99 (training)/>0.99 (validation), enabling rapid predictions (seconds vs. minutes per EM simulation). Radiation characteristics remain suitable (gain ~2–3.2 dBi lower/~3.5–4.3 dBi upper; efficiency >80–85%). The hybrid approach offers scalable, efficient methodology for next-generation dual-band antennas, with novelty in tunable band-prioritized FOM + ANN for legacy monopole enhancement.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1802007</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1802007</link>
        <title><![CDATA[Exploring English for academic purposes learners’ engagement with ChatGPT-4 in academic writing revision: a case study from a private language institute in China]]></title>
        <pubdate>2026-06-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jinming Lawrence Du</author><author>Yiran Chen</author>
        <description><![CDATA[While a growing body of research has examined the affordances and effects of generative artificial intelligence (GenAI) in language learning, less attention has been paid to how learners engage with these tools during the writing revision process across behavioural, cognitive, and affective dimensions. This qualitative multiple-case study investigates how eight English for Academic Purposes (EAP) learners engage with ChatGPT-4.0 while revising academic writing. Data included screen recordings of learner–AI interactions, interaction transcripts, stimulated recall sessions, and semi-structured interviews. The findings reveal that learner engagement is heterogeneous and multidimensional. Behavioural and cognitive engagement ranged from deliberate, evaluative processing to more mechanical uptake, including instances of overreliance on AI-generated feedback. Affective engagement was similarly mixed: while participants valued the immediacy and accessibility of feedback, some reported a lack of social and contextual support compared to teacher guidance. These findings highlight the importance of teacher mediation and critical AI literacy in supporting meaningful engagement with GenAI. The study contributes to emerging research on AI-assisted language learning by specifying the conditions under which GenAI feedback may foster, rather than undermine, reflective and critical engagement in EAP writing.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1834667</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1834667</link>
        <title><![CDATA[Balancing beliefs: exploring preservice teachers’ generative artificial intelligence perceptions and intentions for practice]]></title>
        <pubdate>2026-06-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Mike Karlin</author><author>Cristina Stephany</author><author>'Alohilani Okamura</author><author>Yoon Jin Nam-Huh</author>
        <description><![CDATA[Generative artificial intelligence (GenAI) tools such as ChatGPT have rapidly entered educational contexts, raising questions about their affordances, limitations, and implications for teacher preparation. This qualitative, exploratory case study investigated preservice teachers’ perceptions of ChatGPT across two sites in Los Angeles and Hawai‘i, with 74 participants. Data were collected through surveys embedded in interactive learning modules and analyzed using thematic analysis. Findings indicated alignment in perceived affordances for both students and teachers, including idea generation, and the importance of knowing how to use tools properly. Participants also emphasized limitations such as risks of dependency, plagiarism, and exposure to bias or misinformation. While most preservice teachers expressed willingness to use ChatGPT in their future teaching, few supported future student use. Across responses, participants highlighted the need for training to ensure ethical and effective implementation. This study underscores the importance of integrating both technical and critical AI literacy into teacher education.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1772844</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1772844</link>
        <title><![CDATA[Investigating annotator bias in comment quality and incivility classification by formal education]]></title>
        <pubdate>2026-06-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Lena Wilms</author><author>Anke Stoll</author><author>Marc Ziegele</author>
        <description><![CDATA[Machine learning models used in algorithmic content moderation in online discussions play an increasingly important role in detecting high-quality (e.g., constructive or engaging) and uncivil (e.g., toxic, hateful, or abusive) comments. Trained on human decisions, these models risk reproducing annotator bias, particularly when the perception of such nuanced categories is influenced by social backgrounds and training data lacks sufficient representational diversity. This study investigates the potential impact of annotators’ formal education on the classification of high-quality and uncivil user comments in online discussions. Using a dataset of 13,677 German user comments annotated by crowd workers with low, medium, and high formal education, we investigate divergences in classification outcomes when classifiers are trained on data annotated by different educational groups. In addition, we assessed statistical associations between fine-grained comment features and the annotations provided by each subgroup. Our results indicate that classifiers for high-quality and uncivil comments yield significantly different results when trained on data labeled by annotators with varying educational backgrounds. Furthermore, comment features used in traditional operationalizations of deliberative quality (e.g., solution proposals, additional knowledge) and incivility (e.g., vulgarity, accusation of lying) are more strongly associated with annotations by crowd annotators with medium and high formal education. We argue that such group-specific differences should be considered when developing machine learning models in both practice and research to reflect a more inclusive understanding of valuable and norm-violating user contributions to online discussions.]]></description>
      </item>
      </channel>
    </rss>