Artificial intelligence is increasingly embedded in day-to-day operations, including highly regulated and sensitive sectors, such as autonomous systems, legal practice, healthcare, and cyber-physical environments. As with any technological evolution, its adoption offers opportunities for operational efficiency and effectiveness but also brings significant, complex, and stealthy cybersecurity concerns. Cybersecurity audit readiness has, therefore, become essential for responsible and secure deployment that balances the enhancements of artificial intelligence against sound cybersecurity readiness, particularly as no area is exempt from this wave. This study systematically assesses the cybersecurity audit-readiness evidence reported in publications on systems enabled by artificial intelligence. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework, 22 studies were selected and assessed using a structured Cybersecurity Audit Readiness Algorithm covering eight assessment dimensions. Nine studies (40.9%) reported evidence classified as High readiness, nine (40.9%) as Moderate readiness, and four (18.2%) as Low readiness; these classifications reflect the strength of documented evidence rather than independently verified operational maturity. The synthesis identifies recurring challenges concerning governance integration, audit traceability, assurance measurement, and continuous monitoring. In response, the study proposes a structured research agenda for advancing the design, evaluation, and implementation of measurable, auditable, and cyber-resilient artificial intelligence systems.
Programmable metasurfaces provide transformative platforms for wavefront manipulation and beamforming. However, current works still highly rely on manual control (e.g., pre-specified look-up table rules), which limits their applications for intelligent wireless communications. Here, we propose a large language model (LLM)-assisted pattern-switching technology based on an amplitude–phase-controllable metasurface, which can reason the working states of the metasurface according to semantics and realize the expected beam-steering and beamforming functions. The proposed meta-atom integrates two types of diodes (PIN and varactor diodes), providing independent amplitude–phase control and 360° full-phase coverage. The LLM can conduct the dimensionality reduction in high-level semantics into structured electromagnetic targets and map them to the metasurfaces, controlling the sequences of the metasurface and thereby realizing the desired pattern-switching function. The autonomy refers to task-level autonomous execution after a high-level user requirement is specified, including automatic state searching, threshold-based performance evaluation, and iterative reconfiguration. To verify the feasibility of the proposed strategy, we fabricated an amplitude–phase-controllable metasurface and conducted LLM-assisted functional measurements. The measured results agree well with the expected ones. The proposed method holds great potential for future human–machine interfaces in wireless communications.
Interdisciplinary research at the intersection of reinforcement learning and nonlinear control encompasses a wide range of dynamical systems, closed-loop objectives, and control mechanisms. However, existing reviews usually organise the literature by algorithm category or application scenario, making it difficult to explain both the behavior ultimately produced by the system and the logic through which the controller achieves its objectives. This scoping review aims to map the distribution and thematic structure of the literature on reinforcement learning-based control of nonlinear dynamical systems, from 2021 to 2026. It summarizes the desired closed-loop behaviors and control-design mechanisms and analyses their co-occurrence relationships. The Web of Science Core Collection, Scopus, IEEE Xplore, and five conference proceedings were searched, with eligible sources limited to English-language journal articles and conference papers. After deduplication, 3134 records were retained and rated for analytical relevance on an A–D scale. The A + B-rated corpus (n = 2211) was used for literature-distribution mapping and TF–IDF, truncated singular value decomposition, and Ward clustering analyses with sensitivity tests, while the A-rated core corpus (n = 1187) underwent manual dual-axis coding. The results show that value-function- and optimality-driven mechanisms occur across most closed-loop behavior categories, whereas orbital and motion-reproduction behaviours, synchronization and relational manifolds, and safety-set preservation primarily co-occur with data-driven policy learning, interaction and resource awareness, and safety, robustness, and reliability mechanisms, respectively. Overall, the reviewed literature remains dominated by regulation and tracking and exhibits a many-to-many structure in which value-function- and optimality-driven mechanisms provide a common foundation, while other mechanisms are combined according to the closed-loop objective. The main contribution of this review is a dual-axis analytical framework that links desired closed-loop behaviors to control-design mechanisms. This framework moves beyond the traditional organisation of evidence solely by algorithm category and reveals the structural interdependence between control objectives and design mechanisms. Future research should further investigate transitions between closed-loop behaviors and the integrated design of policy learning with stability, safety, resource constraints, and fault handling.
Background: Pentoxifylline, a hemorheological drug with anti-inflammatory properties, has been proposed as adjunctive therapy in acute myocardial infarction, yet its effect on mortality along with any differences in patient phenotypes remain unclear. Methods: We carried out a registry-based study on 1700 patients with acute myocardial infarction, of which 1684 had a known exposure status to pentoxifylline. The admission-time confounders were analyzed using 1:1 propensity-score matching (PSM) to balance baseline characteristics between pentoxifylline users and non-users. We employed an extreme gradient boosting (XGBoost) model on a hold-out set along with explanations from SHapley Additive exPlanations and Accumulated Local Effects analyses. Results: In the PSM-matched pair analysis, mortality was observed to be 16.4% with and 15.4% without pentoxifylline (risk difference: 1.1% [95% CI: −5.2, 7.4]; odds ratio: 1.09% [95% CI: 0.68, 1.75]. Sensitivity analyses were concordant with the overall findings. Neither any subgroup interaction nor heterogeneity was observed with causal forest. The XGBoost model demonstrated good discriminatory performance (test-set AUC: 0.834, 95% CI: 0.785–0.884). Conclusions: In the present analysis, pentoxifylline use in the intensive care unit was not associated with a reduction in the in-hospital mortality both in overall or in any of the inflammatory, metabolic or infarct-defined subgroup.
As deepfake generation methods are constantly evolving, there is a need for new methods that can detect subtle manipulation artefacts in video content, both in terms of spatial and temporal characteristics. In this study, an encoder–decoder architecture is proposed, where the classification and forensic reconstruction are performed simultaneously at the encoder and decoder side, respectively. The architecture consists of a six-layer convolutional encoder to obtain spatial representations from uniformly sampled facial frames of 128 × 128, and a two-layer bidirectional gated recurrent unit (BiGRU) to model temporal dependency within the sequence of eight facial frames. An attention-pooling mechanism additionally highlights the most instructive temporal characteristics and yields a brief sequence-level representation for binary categorization. In parallel, a forensic decoder reconstructs 64 × 64 Laplacian residual maps, encouraging the network to learn edge-level and texture-level manipulation traces. The framework is trained end-to-end using a weighted combination of cross-entropy classification loss and L1 reconstruction loss. Experiments on the Celeb-DF dataset, comprising 1168 cropped face videos, achieved 93% training accuracy, 90% frame-level classification accuracy, and an ROC-AUC of 0.92. Complementary analysis across spatial, frequency, and temporal feature domains showed that spatial texture characteristics provided the clearest class separation, while the forensic reconstruction branch achieved an MSE of 0.010. Overall, the results demonstrate the value of jointly integrating spatial encoding, temporal sequence modelling, attention-based aggregation, and reconstruction-guided forensic learning for deepfake detection.
