<?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>2026-05-13T18:27:55.783+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1791624</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1791624</link>
        <title><![CDATA[An LLM-based methodology for the automatic detection of bias in the DuoWikiBias corpus]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Karla Salas-Jimenez</author><author>Sergio-Luis Ojeda-Trueba</author><author>Gemma Bel-Enguix</author><author>Edgar Lee-Romero</author><author>Francisco F. López-Ponce</author>
        <description><![CDATA[Bias detection remains a challenge in Natural Language Processing, particularly in non-English contexts, due to the conceptual ambiguity of bias and the scarcity of annotated resources. This study addresses the lack of Spanish-language resources by investigating the automatic detection of framing, epistemological, and demographic1 biases. We introduce DuoWikiBias, a novel parallel corpus derived from Wikipedia for Spanish bias classification. We evaluate Large Language Models (Llama and Gemma) using advanced prompting techniques—CARP and Metacognition—combined with a Gradient Ascent unlearning method to refine model attention. Their performance is compared against classical approaches, including logistic regression with S-BERT embeddings and linguistic features. Results show that advanced prompting substantially improves performance over simple instructions, while the best overall performance (F1 = 0.796) is achieved by combining CARP-based features with Gradient Ascent and a Support Vector Machine classifier. These findings suggest that LLMs are effective for bias-aware representation learning, but hybrid approaches with traditional classifiers remain competitive. This work provides both a validated dataset and a methodological framework for bias detection in Spanish NLP.2]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1747663</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1747663</link>
        <title><![CDATA[Harnessing discrete choice experiments to elicit preferred configurations of trustworthy AI augmented decision support systems for certified crop advisors]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Asim Zia</author><author>Maaz Gardezi</author><author>Xinjing Yu</author><author>Benjamin Ryan</author><author>Scott Merrill</author><author>Eric Clark</author><author>Ali Dadkhah</author><author>Donna M. Rizzo</author><author>John McMaine</author><author>David Clay</author>
        <description><![CDATA[IntroductionThe increase in Artificial Intelligence (AI) and sensor data driven Precision Agriculture (PA) technologies show promise to improve efficiencies in agricultural production systems and decrease adverse impacts of agriculture on environment compared to traditional approaches. Yet, complex trade-offs (e.g. cost, accuracy, precision and data ownership) in the design and configuration of trustworthy AI augmented decision support systems (AI-DSS) for advancing responsible and ethical PA have surfaced. This study harnesses Discrete Choice Experiments (DCEs) to elicit stated preferences of Certified Crop Advisors (CCAs) for informing the design and configurations of trustworthy AI-DSS. The research is guided by two questions and eight associated hypotheses: (a) How do cost, accuracy, precision, and data ownership influence the preferences of CCAs for adopting AI-DSS in agriculture? (b) Which AI perceptions, PA technology concerns and prior DSS experience predict the adoption of AI-DSS configurations?.MethodsSix focus groups informed the design of the choice set, comparing low, medium and high cost AI-DSS with varying accuracy, precision and data ownership attributes. The survey was circulated by Crop Science Society of America to ~2600 CCAs with a lottery-based incentive, leading to 771 responses (response rate = 29.65%). The DCE data were analyzed using a Standard (McFadden) Logit Model, and a Random Utility Mixed Logit Model.ResultsAnalysis showed 25.54% of the participants opted out, and 45.36%, 19.23%, 9.85% prefer low, medium and high-cost AI-DSS, respectively. Marginal improvement of 1% accuracy leads to ~4% (p < 0.001) and spatial precision leads to ~ 1.5% (p < 0.01) increment in the likelihood of AI-DSS adoption.DiscussionAI perceptions, PA technology concerns and prior DSS experience significantly predict the variability in the adoption of three types of AI-DSS.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1759740</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1759740</link>
        <title><![CDATA[A multi-class defect detection method for substations based on the improved YOLOv10n]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Long Huang</author><author>Kangning Li</author><author>Tianren Fu</author><author>Jinlei Zhu</author><author>Yunhao Zhong</author><author>Xingfei Wang</author>
        <description><![CDATA[IntroductionEnsuring the stable operation of power substations is critical for maintaining the reliability of the electrical grid. However, automated inspection of substation equipment remains challenging because multi-class defects are often small, visually blurred, and located in complex backgroundsMethodsTo improve the localization accuracy of small defects with fuzzy features, this paper proposes YOLO-SMALLNET, an improved defect detection algorithm based on YOLOv10n. First, a Detail Information Extraction Convolution module is used to replace the strided convolution modules in the baseline network to preserve fine-grained information during downsampling. Second, a low-level feature fusion detection layer is introduced to reduce small-target feature loss. Third, a Weighted Hybrid Fusion Pyramid Network is adopted to optimize multi-scale feature integration. Finally, a Content-Guided Attention mechanism is integrated to enhance critical defect information while suppressing background noise.ResultsExperimental results show that, compared with the baseline model, YOLO-SMALLNET improves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 by 7.3, 8.2, 3.9, and 3.3%, respectively.DiscussionThe proposed method effectively reduces false detections and missed detections of small defect regions and is suitable for real-time automated inspection of substation equipment.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1767814</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1767814</link>
        <title><![CDATA[Enabling earlier detection of spinal lesions in CT imaging with artificial intelligence—a case study]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Community Case Study</category>
        <author>Marlene Fritzsche</author><author>Patrick Kara-Schmidt</author><author>Matthias Kirchler</author><author>Kenneth Schroeder</author><author>Christian Wiedemeyer</author><author>Lea-Elena Braunschneider</author>
        <description><![CDATA[This study investigates whether an artificial intelligence–based second reader can detect malignant spinal lesions on computed tomography earlier than radiologists, thereby supporting precision oncology through more timely diagnosis and treatment planning. The spine is one of the most common locations for metastases in advanced cancer, significantly influencing symptoms, staging and therapeutic decisions. Delayed or missed detection can, however, impair outcomes. A three-dimensional nnU-Net segmentation model trained on 653 scans was applied to a retrospective cohort of 200 patients who later received confirmed diagnoses of malignant spinal lesions; earlier examinations without documented disease were re-evaluated by one board-certified radiologist both unaided and with AI support. The primary measure was the proportion of malignant spinal lesions detected by the model before their baseline reporting, with AI-derived lead time as a secondary endpoint. The system identified 12 malignant spinal lesions that remained invisible to radiologists on unaided retrospective review, achieving a mean lead time of 228 days, and highlighted 25 additional malignant spinal lesions that were retrospectively visible but initially unreported. Across the cohort, AI flagged earlier findings in 37 patients. These preliminary results suggest that AI-assisted CT interpretation may have the potential to identify sub-visual or otherwise overlooked malignant spinal lesions at an earlier scan date than standard radiologist reporting. Further prospective studies are needed to determine whether these findings translate into clinical benefits such as improved diagnostic completeness, earlier treatment initiation, and better patient outcomes.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1820375</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1820375</link>
        <title><![CDATA[Building MCP-native hierarchical AI scientist ecosystems: a perspective on scaling multi-agent scientific discovery]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Perspective</category>
        <author>Ling Yue</author><author>Ching-Yun Ko</author><author>Pin-Yu Chen</author><author>Shimin Di</author><author>Shaowu Pan</author>
        <description><![CDATA[Large language models (LLMs) are evolving from chatbots with limited tool-using capabilities to agentic AI systems that can perform deep research, assist in proposing hypotheses, help design experiments, automate data analysis, and draft scientific reports. However, there are currently two bottlenecks limiting LLMs' real-world impact on the broader scientific research community beyond academic demonstrations: lack of interoperability (repetitive manual tool-integration is required across scenarios) and the need for scalable coordination (unstructured communication and memory become brittle as the number of agents grows). In this Perspective, we argue that the next phase of agentic scientific discovery requires the development of an ecosystem of protocol-native agents and tools organized through hierarchies inspired by human society, beyond the current paradigm of a single monolithic “AI scientist”. We use Model Context Protocol (MCP) as a concrete example of an emerging interoperability layer for scientific tool and context exchange, and we propose three complementary pathways to increase the scaling capabilities of an MCP-native scientific ecosystem by addressing the composability issues: (1) MCP servers for high-value scientific tools maintained by domain experts, (2) automated transformation of existing code repositories into MCP services, and (3) autonomous invention and evolution of new agents and workflows. Finally, we provide a practical roadmap for scaling AI-driven scientific discovery by expanding tool supply and coordination in MCP-native scientific ecosystems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1716108</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1716108</link>
        <title><![CDATA[Between innovation and risk: artificial intelligence and data protection in digital Mexico]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Emilio J. Medrano-Sánchez</author><author>Elizabeth Ruiz-Ramírez</author><author>Mariela L. Ayllon</author>
        <description><![CDATA[At a global level, the rapid adoption of artificial intelligence (AI) brings significant risks to data privacy, prompting the development of legal frameworks. In Latin America, including Mexico, where such frameworks remain emergent, the issue gains particular relevance. This study analyzed how perceptions of AI are associated with perceptions of personal data protection among university-educated professionals residing in Mexico City in 2025, based on privacy theory, the Technology Acceptance Model (TAM), and their extensions. A quantitative, non-experimental, correlational, and cross-sectional approach was employed. Data were collected from 101 university-educated professional participants residing in Mexico City using an expert-validated, 24-item Likert-type questionnaire. Data analysis was conducted in SPSS using non-parametric correlation tests (Spearman and Kendall). Results indicated that a more favorable perception of AI was associated with more favorable perceptions of personal data protection and its dimensions. The correlations demonstrated a moderate and significant positive association (p < 0.001) between perceptions of AI and overall perceptions of data protection. Furthermore, significant positive correlations were found across each evaluated dimension, thus confirming the four specific hypotheses. From a theoretical perspective, the findings suggest that contextual factors modulate the AI-privacy relationship, contributing a Latin American perspective to the literature. On a practical and social level, recommendations aligned with SDG 16 include strengthening institutional frameworks (regulation, transparency, digital education), fostering public-private collaboration, and promoting digital oversight and literacy to achieve informed trust.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1784986</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1784986</link>
        <title><![CDATA[A feasibility study on integrating Generative AI and video conferencing into Indonesian recruitment processes]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Muhammad Ramaditya</author><author>Mohammad Ikhsan</author><author>Muhammad Firdaus Syawaludin Lubis</author><author>Corina D. Riantoputra</author>
        <description><![CDATA[The recruitment landscape in Indonesia faces a critical efficiency gap, with a global average time-to-hire of 44 days and a significant mismatch between labor supply and demand. This study evaluates “The Platform,” an innovative recruitment tool integrating Generative Artificial Intelligence (Gen-AI) and video conferencing to modernize interviews. To assess the project’s viability, we used a strategic management approach, including Political, Economic, Social, Technological, Environmental, and Legal (PESTEL) analysis to evaluate macro-environmental risks and opportunities; Porter’s Five Forces to analyze the competitive landscape and industrial profitability; and Strength, Weakness, Opportunity, and Threat (SWOT) analysis to identify internal strengths and weaknesses in the context of external factors. Furthermore, data from Focus Group Discussions (FGDs) with 58 Human Resources Managers were analyzed using a thematic approach to identify, assess, and map emerging Artificial Intelligence (AI) assumptions and user requirements. Findings reveal a green light for adoption, supported by high digital literacy in Indonesia, where a Serviceable Obtainable Market (SOM) is estimated at US$1.75 billion. Key success factors identified include the platform’s ability to conduct interviews in the local language and its alignment with national digital transformation goals. The study identifies culture lag and linguistic nuances as primary barriers to efficiency. Consequently, the study concludes that by addressing these identified efficiency gaps through robust data security and localized AI models, the platform is highly feasible and poised to revolutionize the Indonesian HR ecosystem.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1783410</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1783410</link>
        <title><![CDATA[How to systematically and quantifiably remove meaning?]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Frida Proschinger Åström</author><author>Arend Hintze</author>
        <description><![CDATA[Large language models increasingly mediate real-world tasks, yet we lack systematic ways to quantify how their performance degrades when the meaning of their inputs is eroded. To bridge this gap, we developed a framework to semantically erode meaning and quantify its intensity, grounded in discourse analysis, psycholinguistics, and software engineering, comprising five theoretically motivated methods: omission of key information and context, lexical substitution with near-synonyms, increased abstraction, structural obfuscation and renaming, and injection of logical errors. We applied these erosion operators across five domains and quantified their effects on model performance using a publicly available language model. A two-way Analysis of Variance (ANOVA) revealed significant main effects of both domain and erosion method, as well as a significant interaction, indicating that the impact of semantic degradation depends jointly on how text is eroded and how domain-specific information is encoded. Logical error erosions proved especially damaging for code generation, whereas structural obfuscation most strongly impaired news and instruction tasks. Epistasis analysis of pairwise erosion unions showed that some combinations produced super-additive degradation while others exhibited compensatory effects. These domain-by-erosion profiles provide diagnostic insight into where multi-step large language model (LLM) pipelines are most likely to fail and suggest that robustness benchmarks should probe models along domain-specific vulnerability dimensions rather than relying on generic perturbations. Semantic erosion thus offers a principled tool for turning model failure into evidence about how language models structure and degrade meaning.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1784484</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1784484</link>
        <title><![CDATA[Adversarial robustness of LLM-based multi-agent systems for engineering problems]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Lorenz Wiesmeier</author><author>Matthias Busch</author><author>Marius Tacke</author><author>Kevin Linka</author><author>Christian Cyron</author><author>Roland Aydin</author>
        <description><![CDATA[Large language models (LLMs) are increasingly deployed in multi-agent systems (MAS), including for solving engineering problems. Unlike purely linguistic tasks, engineering workflows demand formal rigor and numerical accuracy, meaning that adversarial perturbations can cause not just degraded performance but systematically incorrect or unsafe results. In this work, we present one of the first systematic studies of adversarial robustness of LLM-based MAS in engineering contexts. Using representative problems—including pipe pressure loss (Darcy-Weisbach), beam deflection, mathematical modeling, and graph traversal—we investigate how misleading agents affect collaborative reasoning and quantify error propagation under controlled adversarial influence. Our results show that adversarial vulnerabilities in engineering differ from those observed in generic MAS evaluations in important aspects: system robustness is sensitive to task type, the subtlety of injected errors, and communication order among agents. In particular, engineering tasks with higher structural complexity or easily confusable numerical variations are especially prone to adversarial influence. We further identify design choices, such as prompt framing, agent role assignment, and discussion order, that significantly improve resilience. These findings highlight the need for domain-specific evaluation of adversarial robustness and provide actionable insights for designing MAS that are trustworthy and safe in engineering applications.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1813511</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1813511</link>
        <title><![CDATA[EcoStack-Pro: an adaptive federated learning framework for interpretable ESG auditing across heterogeneous industrial sectors]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Md. Abul Kalam Azad</author><author>Abdul Kadar Muhammad Masum</author><author>Md. Abdur Rahman</author><author>Md. Tofael Ahmed Bhuiyan</author><author>Farzan Majeed Noori</author><author>Md Zia Uddin</author>
        <description><![CDATA[IntroductionThe paradigm shifts toward environmental, social, and governance (ESG) metrics has necessitated advanced auditing systems capable of analyzing complex, non-financial performance indicators. However, traditional centralized artificial intelligence (AI) models conflict with increasingly stringent data privacy regulations, while conventional federated learning approaches struggle to converge under the high statistical heterogeneity and data imbalance typical of diverse industrial sectors.MethodsTo address the trade-off between high-precision forecasting and data sovereignty, this study proposes EcoStack-Pro, a decentralized auditing framework driven by a stacked ensemble of LightGBM, XGBoost, and Gradient Boosting regressors, optimized via a Bayesian ridge meta-learner. Central to this architecture is the Fed-GenAdaptive algorithm, which employs a soft-gating mechanism with softmax normalization to dynamically weight client contributions according to their local validation errors and generalization gaps.ResultsUtilizing a stratified dataset of 21,400 firm-year observations across 10 distinct industrial clients, the framework achieves a test-set R2 of 0.9614. This performance retains 98.2% of the predictive power of the centralized upper bound (R2 of 0.9790) while strictly preserving corporate privacy.DiscussionFurthermore, the integration of Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) enhances model interpretability, elucidating the non-linear drivers of governance ratings. These results demonstrate that adaptive, diverse ensemble strategies can overcome the limitations of single-model federated baselines, providing a robust framework for secure, cross-sector sustainable finance auditing.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1774453</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1774453</link>
        <title><![CDATA[Predictive models of suicidal ideation risk in perinatal-stage women based on sociodemographic and clinical data]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>José Manuel Martínez-Ramírez</author><author>Rocío Adriana Peinado-Molina</author><author>Antonio Hernández-Martínez</author><author>Juan Miguel Martínez-Galiano</author>
        <description><![CDATA[IntroductionSuicidal ideation in women during the perinatal period has become a growing public health problem, with a prevalence ranging from 8 to 19%. Its etiology is multifactorial and carries additional consequences for both the newborn and the woman beyond death itself.AimThis study aims to predict the risk of suicidal ideation in perinatal women by using artificial intelligence models based on sociodemographic and clinical data.MethodsAn analytical observational study was conducted with a sample of 908 Spanish women during the perinatal period, collecting relevant data. To predict the risk of suicidal ideation, five machine learning models (OneR, JRIP, FURIA, J48 and Random Forest) were employed, as they provide rules or trees that can be easily followed to gather additional information. The metrics used to evaluate the performance of the models included accuracy, precision, recall and F1-score.ResultsThe models show an accuracy of around 60% in most cases. The model that performs the worst is OneR, with an accuracy of less than 50%. The Random Forest model stood out for its higher accuracy. The metrics of this model (Random Forest) were Accuracy (%) 0.639 ± 0.05, Precision: 0.634, Recall: 0.634 F1:0.634 and AUPRC: 0.632. Factors identified as predictors of suicidal ideation risk included low birth weight, history of mental health problems, problems of intimate partner violence, low income and smoking.ConclusionIn conclusion, predictive models based on sociodemographic data and clinical variables show a moderate ability to predict suicidal ideation risk in perinatal women.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1805539</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1805539</link>
        <title><![CDATA[Artificial intelligence agents as advanced decision support systems in public decision-making: evidence from Peru]]></title>
        <pubdate>2026-05-13T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Emilio J. Medrano-Sánchez</author><author>Jadira Jara</author>
        <description><![CDATA[Decision making in the public sector faces persistent challenges that have driven the analysis of technologies aimed at strengthening institutional performance. Within this context, the present study examined the association between the valuation of artificial intelligence agents (AIA) and decision making among public officials with national-level responsibilities in Peru. In this study, AIA are conceptualized as advanced decision support systems (DSS) that complement human judgment in complex public-sector decisions. A Likert-scale survey was administered to 93 participants and showed high internal consistency (Cronbach's alpha = 0.952). Given the non-parametric nature of the data, associations were estimated using Kendall's Tau-b and Spearman's Rho. The results revealed a positive and statistically significant association between the valuation of AIA and overall decision-making (Tau-b = 0.588; Rho = 0.653; p < 0.001). Complementarily, the observed pattern was consistent with the notion that higher valuation of AIA is linked to more favorable perceptions of the decision-making process across the analyzed dimensions. In conclusion, the findings support that, in the Peruvian context, the valuation of AIA is associated with decision making perceived as more effective, suggesting their relevance as a complementary support to human judgment and as an AI-based DSS to guide evidence-based modernization of public management.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1811692</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1811692</link>
        <title><![CDATA[Advanced behavioral malware detection: a comprehensive MLOps framework with federated learning and real-time drift detection]]></title>
        <pubdate>2026-05-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Mohammed El-Hajj</author><author>Mohammad Al Jawad Zeineddine</author>
        <description><![CDATA[This paper presents a comprehensive MLOps framework for behavioral malware detection that addresses critical challenges in generalization, collaboration, and operational resilience. We introduce three methodological contributions: (1) a formalized Leave-One-Experiment-Out (LOEO) validation protocol that provides conservative assessment of generalization to novel attack methodologies, revealing a 12.3% accuracy drop compared to conventional evaluation; (2) a domain-optimized feature engineering pipeline that transforms raw process telemetry into hierarchical behavioral signatures while maintaining 99.2% accuracy with 50% reduced inference latency; and (3) a hybrid federated learning architecture enabling privacy-preserving collaboration with 75.1% accuracy while maintaining (ϵ, δ)-differential privacy guarantees (ϵ = 3.2, δ = 10−5). A real-time drift detection engine with sub-500 ms latency identifies concept drift using ensemble detection and triggers automated retraining with total recovery time < 5 min (mean 4.2 min). Comprehensive evaluation across 2.74 million behavioral samples from 104 distinct malware experiments validates our approach using up to 104 federated clients, achieving 10,000+ events/s throughput in simulated environments. Architectural projections based on hierarchical aggregation suggest potential scalability to 5,000+ clients, though this remains unvalidated future work. This work bridges the gap between academic research and operational cybersecurity requirements through a production-oriented MLOps implementation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1797906</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1797906</link>
        <title><![CDATA[Short-term load forecasting using a metaheuristic optimized temporal fusion transformer with decomposition technique]]></title>
        <pubdate>2026-05-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Radhika Chandrasekaran</author><author>Senthil Kumar Paramasivan</author>
        <description><![CDATA[Short-term load forecasting plays a vital role in today's modern life to ensure the balance between energy demand and supply. Dynamic variations in weather and electricity consumption patterns can significantly influence load patterns, resulting in complex modeling and challenging forecasting accuracy due to non-stationary, non-linear patterns. Traditional statistical methods are simple and interpretable; however, they struggle to capture temporal and non-linear dependencies. Although machine learning can address these challenges, it struggles with long-range temporal dependency. In contrast, deep learning models can automatically extract temporal patterns and relevant features from raw data. This study proposes a Temporal Fusion Transformer (TFT) based on Multivariate Variational Mode Decomposition (MVMD) and optimized with the GOAT Optimization Algorithm (GOA). The deep architecture of the transformer model is considered an alternative owing to its self-attention mechanism, its efficiency in capturing long sequences, its parallel processing capabilities, and its interpretability through attention weights, making it suitable for multivariate short-term load forecasting. The significant benefit of combining MVMD with TFT is that it enhances feature extraction by decomposing complex multivariate time series data into components with different frequencies and reduces noise. The MVMD addresses the non-stationary and non-linear challenges of energy load data with reduced complexity and interpretability. Furthermore, the GOAT Optimization Algorithm was incorporated for hyperparameter tuning of the TFT model to enhance its performance. The error in the model is evaluated using mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and symmetric mean absolute percentage error (sMAPE). The proposed model outperforms other comparison models. Finally, the model is interpreted using SHapley Additive exPlanation (SHAP) analysis to understand the impact of features on the model's prediction.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1760246</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1760246</link>
        <title><![CDATA[Language-based personality assessment from life narratives: a focus on model interpretability and efficiency]]></title>
        <pubdate>2026-05-08T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Rasiq Hussain</author><author>Zerui Ma</author><author>Ritik Khandelwal</author><author>Joshua Oltmanns</author><author>Mehak Gupta</author>
        <description><![CDATA[Natural Language Processing (NLP) enables novel approaches to personality assessment by analyzing rich, open-ended narratives, rather than relying solely on structured questionnaires. While most existing work focuses on short-form social media texts, this study shifts attention to predicting the Big Five personality traits, a framework closely tied to mental-health outcomes, from long-form life narrative interviews. Each narrative exceeds 2,000 words, posing significant challenges for standard language models. We propose a two-step modeling framework that captures contextual representations of long texts while prioritizing interpretability and computational efficiency. First, we extract contextual embeddings using a sliding-window finetuning strategy on pretrained transformer models. These embeddings are then processed with Recurrent Neural Networks (RNNs) equipped with attention mechanisms to capture long-range dependencies and provide interpretable outputs. Our hybrid approach effectively combines the representational power of transformers with the sequence-handling strengths of RNNs. Through comparisons with state-of-the-art long-context models and interpretability analyses, we demonstrate improvements in prediction accuracy, computational efficiency, and interpretability. These findings underscore the potential of interpretable and efficient language models for assessing personality from life narratives.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1819009</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1819009</link>
        <title><![CDATA[Learning instance-specific counterfactual models for continuous treatments using hypernetworks]]></title>
        <pubdate>2026-05-08T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Roger Pros</author><author>Jordi Vitrià</author>
        <description><![CDATA[In recent years, non-linear machine learning techniques have attracted increasing attention for estimating treatment effects from observational data. While most existing methods focus on binary treatment scenarios, the estimation of treatment effects for continuous interventions remains a critical challenge in many real-world applications. In this work, we introduce a novel neural network-based approach for estimating continuous treatment effects by leveraging hypernetworks to model counterfactual outcomes across treatment levels. This approach extends the principles of binary treatment effects computation to the continuous domain, addressing the key challenge of treatment relevance. By generating weights for a fixed network that predicts potential outcomes, our architecture ensures that the treatment variable retains its causal significance while maintaining the flexibility of deep learning models. Through extensive experiments on synthetic and semi-synthetic datasets, we demonstrate that our approach outperforms existing methods in terms of precision. The results highlight the advantages of explicitly modeling the relationship between treatment levels and outcomes, particularly in settings where traditional methods struggle with high-dimensional confounders or non-linear treatment-response dynamics.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1817837</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1817837</link>
        <title><![CDATA[Spark: modular spiking neural networks]]></title>
        <pubdate>2026-05-08T00:00:00Z</pubdate>
        <category>Technology and Code</category>
        <author>Mario Franco</author><author>Carlos Gershenson</author>
        <description><![CDATA[Nowadays, neural networks act as a synonym for artificial intelligence. Present neural network models, although remarkably powerful, are inefficient both in terms of data and energy. Several alternative forms of neural networks have been proposed to address some of these problems. Specifically, spiking neural networks are suitable for efficient hardware implementations. However, effective learning algorithms for spiking networks remain elusive, although it is suspected that effective plasticity mechanisms could alleviate the problem of data efficiency. Here, we present a new framework for spiking neural networks—Spark (https://github.com/Nogarx/Spark)—built upon the idea of modular design, from simple components to entire models. The aim of this framework is to provide an efficient and streamlined pipeline for spiking neural networks. We showcase this framework by solving the sparse-reward cartpole problem with simple plasticity mechanisms. We hope that a framework compatible with traditional ML pipelines may accelerate research in the area, specifically for continuous and unbatched learning, akin to the one animals exhibit.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1785979</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1785979</link>
        <title><![CDATA[AI-driven skill volatility and the emergence of re-skilling fatigue: the human cost of perpetual learning]]></title>
        <pubdate>2026-05-08T00:00:00Z</pubdate>
        <category>Perspective</category>
        <author>Tittu Elizabath Biju</author><author>Anusree Ambady</author><author>Thomas K V</author>
        <description><![CDATA[Artificial intelligence (AI) is rapidly reshaping knowledge-intensive work by automating, augmenting, and reconfiguring core professional activities. While continuous reskilling is widely promoted as a solution to AI-driven disruption, little attention has been paid to its cumulative psychological costs. This paper introduces the concept of reskilling fatigue to explain the human consequences of persistent skill volatility among Established Knowledge Professionals (EKPs), mid-career professionals whose roles, identities, and value are grounded in accumulated expertise and professional judgment. Based on Job Demands–Resources (JD-R) theory and Conservation of Resources (COR) theory, the paper conceptualizes an AI-induced reskilling loop in which ongoing technological change leads to skill erosion, continuous reskilling demands, cognitive and emotional depletion, and reinforced learning as a defensive response to perceived obsolescence. Unlike restoring stability, this cycle intensifies anxiety, undermines mastery, and erodes professional confidence. As a contribution to theory and practice, the paper advances a set of sustainable, collective strategies such as role-linked learning, protected learning time, skill prioritization, and phased AI adoption to interrupt the reskilling loop and redistribute adaptive demands across organizations. By reframing reskilling as a shared, supported, and bounded process, this paper highlights pathways through which AI-driven change can foster long-term career resilience, professional identity renewal, and sustainable human–AI integration.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1733709</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1733709</link>
        <title><![CDATA[Data-efficient image transformer for landscape character classification and visual comfort prediction in Chinese hospitals]]></title>
        <pubdate>2026-05-08T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yue Li</author><author>Guo Yue</author><author>Riyadh Mundher</author><author>Abdulrahman M. Abdulghani</author>
        <description><![CDATA[In hospital landscapes, where visual comfort influences stress recovery and patient satisfaction, reliable computational tools are needed to link landscape character with human perception. However, existing research on therapeutic landscapes in healthcare has largely focused on qualitative evaluations and design guidelines, with limited development of integrated, interpretable computational models that quantitatively connect landscape characteristics with human perception outcomes. This study addresses this gap by developing an AI-driven decision support system that integrates landscape character classification with visual comfort prediction in Chinese hospital settings. From 488 images collected across three hospitals, 30 representative images were evaluated by 252 respondents. Perception scores were assigned to all images based on landscape character, creating a labeled dataset. A Data-Efficient Image Transformer (DeiT) with dual prediction heads was developed for simultaneous landscape character classification and continuous visual comfort score regression. The model achieved a 96.34% classification accuracy and a mean absolute error (MAE) of 0.055 for visual comfort prediction, substantially outperforming ResNet-50 (accuracy: 89.39%; MAE: 0.148) and the standard Vision Transformer (ViT) (accuracy: 94.06%; MAE: 0.155). The DeiT model demonstrated 19–26% faster convergence and 62–65% improved visual comfort prediction. These results demonstrate that therapeutic landscape qualities, often regarded as subjective, exhibit consistent, computationally learnable patterns. The validated framework provides landscape architects, hospital planners, and administrators with an evidence-based tool for systematic therapeutic landscape evaluation and optimization.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1823559</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1823559</link>
        <title><![CDATA[MCEEGNet: a multi-cue EEG network for quantitative assessment of depression using emotional stimuli-induced EEG signals]]></title>
        <pubdate>2026-05-08T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Xinmin Ding</author><author>Xingyu Chen</author><author>Minpeng Xu</author><author>Yuanmin Zhang</author><author>Hongli Chang</author>
        <description><![CDATA[Depression significantly affects health, manifesting as alterations in typical emotional responses. Its diagnosis depends on subjective evaluations by clinicians, which are often time-intensive. Electroencephalogram (EEG) signals offer a viable solution for aiding diagnosis through computational means. However, current methods primarily focus on binary classification of depression, neglecting the quantification of depression risk. We propose the Multi-Cue EEG Network (MCEEGNet), which consists of parallel branches of EEGNet that extract features from various emotional stimuli to approximate scores on the Patient Health Questionnaire (PHQ-9). MCEEGNet aims to identify depression in patients using EEG signals and assess the severity of the condition. Our method achieved 91.13% accuracy in classification and reported Mean Squared Error (MSE) and Mean Absolute Error (MAE) of 20.45 and 3.28, respectively, in the Multi-Emotion Induced EEG Depression Database. The experimental outcomes suggest that MCEEGNet is highly effective in diagnosing subthreshold depression, offering a comprehensive system for evaluating clinical depression through EEG analysis influenced by multiple emotional cues, thereby meeting the need for quantitative depression evaluation.]]></description>
      </item>
      </channel>
    </rss>