The emergence of ChatGPT in late 2022 has rendered generative discourse models essential to daily living. The increasing user expectations have elevated the necessity to improve these model's capacities to address complex issues in current research. This study examines the influence of Recursive Embedding Fine-Tuning (REFT) on enhancing retrieval accuracy and overall efficacy in biomedical question answering systems. Integrating REFT with the Reason and Action framework (ReAct) and retrieval-augmented generation (RAG) resulted in significant enhancements in the model's capacities to extract information and execute logical reasoning. We evaluated the REFT technique using the PubMed BIOASQ dataset across many reasoning tasks, encompassing both long-form and shortform question answering in English, along with supportive and comparison reasoning contexts. This study highlights the importance of the Reasoning and Action (ReAct) technique in REFT-based models, addressing previous shortcomings in handling long-form question-answering tasks in biological disciplines. Our experimental findings demonstrate significant improvements in performance across key metrics, with the model achieving 97.88% accuracy@10 after minimal training cycles, indicating a 1.1% enhancement in retrieval efficacy. The model exhibited notable learning efficiency, reducing training loss by 62.7% in the early epochs and achieving rapid stability with a final loss of 0.074. Additional metrics, including NDCG@10 (91.59%) and MRR@10 (89.52%), further validate the effectiveness of our approach in enhancing retrieval accuracy and ranking precision in biological question answering systems.

The Probabilistic Neural Pathways (PNP) framework in this study introduces a novel approach to enhancing the performance of large language models (LLMs) through recursive context understanding. By integrating probabilistic reasoning mechanisms, PNP enables LLMs to generate more coherent and contextually relevant text, addressing limitations in traditional models. The mathematical formulation of PNP provides a robust foundation for its implementation, facilitating seamless integration into existing open-source LLM architectures. Experimental evaluations demonstrate significant improvements in language modeling tasks, with notable reductions in perplexity and enhancements in BLEU and ROUGE scores. Qualitative analyses reveal that PNP models maintain thematic consistency over extended passages and accurately interpret idiomatic expressions. However, the integration of PNP introduces additional computational overhead, necessitating further optimization for deployment in resource-constrained environments. Scalability assessments indicate that the effectiveness of PNP amplifies with increasing model size, leveraging greater representational capacity to achieve superior performance. The trade-offs between enhanced capabilities and increased computational requirements highlight the need for strategies to balance performance improvements with computational efficiency. The PNP framework presents a promising advancement in the field of LLMs, offering substantial improvements across multiple natural language processing tasks.

Federated Learning (FL) is a powerful Machine Learning (ML) technique that allows multiple clients to collaborate on training models while keeping their data private. Unlike traditional centralized methods, FL ensures that data are kept separate, which helps to protect privacy. However, an important area that needs more research while using the FL system is detecting harmful models within the Internet of Things (IoT) context. For example, poisoning attacks, where compromised clients introduce harmful data, can degrade the model's overall performance or lead to incorrect predictions. This paper comprehensively reviews of recent attacks in FL within IoT networks, along with defense mechanisms and common FL frameworks. It begins by highlighting the significance of FL in IoT networks, exploring its applications, benefits, and inherent security challenges. It then explores specific attacks targeting FL in IoT networks. The defensive strategies are evaluated, including their performance metrics, datasets used, and related work, providing a comparative analysis of these techniques. Common FL frameworks and their criteria are reviewed. Our goal is to offer a detailed understanding and solutions to enhance the strength and resilience of FL systems in IoT networks.

Vehicle models are essential for advancing intelligent transportation technologies such as automated driving and vehicle stability control functions. To meet real-time and safety requirements, vehicle dynamics must be predicted accurately with low computational effort. However, first-principle models are often associated with incomplete modeling assumptions, unknown parameters, and long computation times during prediction. In contrast, purely data-driven approaches such as artificial neural networks rely on extensive data and can only learn output components for which measured data is available. This work investigates a hybrid approach to address these issues by using physics-informed neural networks (PINNs). We formulate two different PINN architectures, namely a continuous and discrete PINN, for a tractor-semitrailer with partially measured states due to sparse sensor equipment. The proposed PINNs are compared with a gated recurrent unit (GRU) and an identified physics-driven state-space model, as both represent "gold-standard" approaches in their respective modeling worlds. All models are tested for various real-world maneuvers with the tractor-semitrailer. The continuous PINN achieves a 30.4% improved accuracy as well as a 69.4% reduction in computation time compared to the slow physics-driven model. The discrete PINN reaches an accuracy improvement of 28.1%. In contrast to the GRU, the PINNs allow predictions of unmeasured states. Even for the measured states, both PINNs are more accurate than the black-box model.

Antimicrobial resistance is recognized by the World Health Organization as a significant global health threat. The accurate identification of bacterial susceptibility to antibiotics is crucial. However, clinical laboratories often take several days to complete this process, and practitioners rely on probabilistic and empirical reasoning, coupled with local hospital guidelines. In this work, we propose an attention-based bidirectional-Long Short-Term Memory network to predict antibiotic resistance. More precisely, the model is able to give predictions at each stage of the bacterial identification process for a set of 47 antibiotics and combination, in order to support clinicians in their decision process. Excellent results were achieved, with the area under the receiver operating characteristic curve and area under the precision-recall curve reaching up to 0.9, averaged across all antibiotics, at the last stages. The attention mechanism was used to visualize the importance attributed to each features, allowing better interpretation. Additionally, the model is integrated into a user-friendly and responsive web application, accessible on both mobile phones and desktops, to be used as a clinical a decision support system.

The cyclic prefix (CP) is a fundamental feature that enables orthogonal frequency division multiplexing (OFDM) to achieve efficient frequency-domain equalization and resilience against delay spread (DS). However, the trade-off is a reduction in spectral efficiency (SE), prompting the need for methods to minimize CP length without compromising performance. This paper introduces an innovative CP reduction approach leveraging array non-stationarity in extra-large MIMO (XL-MIMO) systems, particularly through the concept of visibility regions (VRs). Unlike conventional methods that adopt a worstcase scenario for CP length, our approach customizes the CP length based on the maximum excess delay of each VRs, thereby reducing the CP overhead. To enhance diversity gain, maximal ratio combining (MRC) is used to optimize the contributions from antenna elements. Simulation results validate our method's ability to improve SE while reducing latency and maintaining low computational complexity. The proposed approach demonstrates practical gains with standard CP configurations, marking a significant step toward more efficient and robust XL-MIMO OFDM systems.

Accurate prediction of cholesterol levels is important for early diagnosis and intervention in cardiovascular health, particularly among women, whose lipid metabolism is profoundly influenced by physiological and hormonal factors, such as estrogen fluctuations during menopause. These hormonal changes contribute to increased risks of dyslipidemia, emphasizing the need for sex-specific predictive models. This study presents an explainable machine learning framework using decision treebased models trained on the Study of Women's Health Across the Nation (SWAN) dataset. By applying Recursive Feature Elimination (RFE), we demonstrated that even with just 10 features, it is possible to achieve an accuracy of 0.74 in detecting high cholesterol levels, significantly reducing the burden of data collection. Through decision tree visualization, we identified the substantial influence of the DIABETE10 feature, which received the highest SHAP value and served as the sole criterion for splitting the tree into high and low cholesterol branches. These findings highlight the importance of diabetes in cholesterol prediction, underscoring its clinical relevance. The integration of Shapley Additive Explanation (SHAP) values and decision path visualization enhances model interpretability, providing actionable insights for personalized healthcare and supporting trust in machine learning applications in high-stakes domains.

Mapping wave field in space has many applications such as optimizing design of radio antennas, improving and developing ultrasound transducers, and planning and monitoring the treatment of tumors using high-intensity focused ultrasound (HIFU). Currently, there are methods that can map wave fields remotely or locally. However, there are limitations with these methods. For example, when mapping the wave fields remotely, the spatial resolution is limited due to a poor diffractionlimited resolution of the receiver, especially when the fnumber of the receiver is large. To map the wave fields locally, the receiver is either subject to damage in hazardous environments (corrosive media, high temperature, and high wave intensity, etc) or difficult to be placed inside an object. To address these limitations, in this paper, the PSF-modulation super-resolution imaging method (January, 2024 IEEE TUFFC) was applied to map pulse ultrasound wave fields remotely at a high spatial resolution, overcoming the diffraction limit of a focused receiver. For example, to map a pulse ultrasound field of a full-width-at-half-maximum (FWHM) beam width of 1.24 mm at the focal distance of a transmitter, the FWHM beam widths of the super-resolution mapping of the pulse wave field with a spherical glass modulator of 0.7-mm diameter at two receiver angles (0 o and 45 o) were about 1.13 mm and 1.22 mm respectively, which were close to the theoretical value of 1.24 mm and were much smaller than the diffraction-limited resolution (1.81 mm) of the receiver. Without using the super-resolution method to remotely map the pulse wave field, the FWHM beam width was about 2.06 mm. For comparison, the FWHM beam width obtained with a broadband (1-20 MHz) and 0.6 mm diameter polyvinylidene fluoride (PVDF) needle hydrophone was about 1.41 mm. In addition to the focused pulse ultrasound wave field, a pulse Bessel beam near the transducer surface was mapped remotely with the super-resolution method, which revealed high spatial frequency components of the beam.

The dynamic and unknown human behaviors in human-robot interaction make it challenging for collision-free robot manipulation. Although sampling-based MPC has achieved real-time control in the above scenarios, it is hard to handle equality hard constraints, such as working along a specified trajectory, due to sampling disturbances. To improve manipulation performance under multiple constraints, this paper presents a novel constrained sampling-based MPC (CSMPC) method using path integral. Firstly, hierarchical optimization combining policy sampling projection and the Lagrange multiplier method is used to handle equality hard constraints for high-precision manipulation tasks. Secondly, collision avoidance and smooth motion are modeled as inequality soft constraints, where collision detection and time series prediction are used to ensure the safety and smoothness of dynamic interaction. Finally, an adaptive noise method is built to improve the stability of physical robot manipulation. The simulation and experiment results demonstrate that the proposed method enables a 7-DOF robot manipulator to achieve precise manipulation while avoiding dynamic obstacles.

The continuous evolution of encryption-based threats has pushed the development of robust and adaptive detection frameworks capable of addressing the complexities introduced by modern attack strategies. In this study, a novel approach leveraging cryptographic execution profiling is proposed, focusing on runtime behavioral telemetry to identify anomalous encryption patterns associated with malicious activities. The method employed entropy analysis, timing metrics, and clustering-based anomaly detection to capture distinctive features of malicious encryption processes, enabling precise differentiation between benign and harmful operations. Experimental evaluations demonstrated high detection accuracy across diverse ransomware families, achieving superior performance compared to traditional signature-based and heuristic approaches. The modular and lightweight architecture of the framework facilitated efficient deployment across heterogeneous environments. Comparative analyses highlighted the framework's capability to address the limitations of machine learning-based detection systems, particularly in terms of data dependency and operational complexity. Resource utilization metrics revealed consistent memory usage and manageable CPU demands. Entropy pattern analyses uncovered irregular but distinct distributions across various encryption schemes, reinforcing the method's effectiveness in isolating ransomware behaviors. The findings contribute a transformative perspective to behavioral detection strategies, showing the importance of telemetry-driven approaches in enhancing the resilience of cybersecurity systems.

Optical polarizers, which selectively transmit light with specific polarization states, are essential components in modern optical systems. Here, we experimentally demonstrate integrated waveguide and microring resonator (MRR) polarizers incorporating reduced graphene oxide (rGO). 2D graphene oxide (GO) films are integrated onto silicon photonic devices with precise control over their thicknesses and sizes, followed by GO reduction via two different methods including uniform thermal reduction and localized photothermal reduction. We measure devices with different lengths, thicknesses, and reduction degrees of the GO films. The results show that the devices with rGO exhibit better performance than those with GO, achieving a polarization-dependent loss of ~47 dB and a polarization extinction ratio of ~16 dB for the hybrid waveguides and MRRs with rGO, respectively. By fitting the experimental results with theory, it is found that rGO exhibits more significant anisotropy in loss, with an anisotropy ratio over 4 times that of GO. In addition, rGO shows higher thermal stability and greater robustness to photothermal reduction than GO. These results highlight the strong potential of rGO films for implementing high-performance polarization selective devices in integrated photonic platforms.

The increasing sophistication of cyberattacks has necessitated the development of innovative detection mechanisms capable of identifying malicious activities with high accuracy and minimal latency. Through the integration of Structural Mutation Analysis and quantum-inspired techniques, a robust framework was proposed to address the challenges posed by modern ransomware variants. The approach leveraged advanced pattern recognition methodologies to capture the structural and behavioral anomalies introduced during ransomware execution, enabling precise classification. Quantum-inspired principles facilitated the identification of complex, non-linear relationships within file trace sequences, enhancing the system's ability to detect previously unseen threats. By employing a combination of reinforcement learning and hybrid machine learning architectures, the detection system demonstrated adaptability and scalability across diverse datasets and attack scenarios. The experimental results highlighted the system's efficiency in handling polymorphic and hybrid encryption schemes, maintaining accuracy rates exceeding 94% and low false-positive rates. A modular system architecture ensured seamless integration with existing cybersecurity infrastructures while supporting real-time operations through efficient computational pipelines. Additionally, empirical analyses revealed that the method provided actionable insights into ransomware propagation dynamics, enabling early-stage mitigation strategies. Advanced visualization techniques further contributed to the interpretability of the detection outcomes, reinforcing its practical applicability.

The Generative Cryptographic Pattern Disruption Analysis (GCPDA) framework offers a novel approach by analyzing cryptographic patterns to identify malicious activities. Through comprehensive evaluations, the framework demonstrated detection accuracies exceeding 95% across various ransomware families, including WannaCry, Locky, CryptoLocker, Petya, and Cerber. Comparative analyses revealed that GCPDA outperforms traditional detection methods, achieving a detection accuracy of 97.0% with a false positive rate of 2.5%. The framework maintained low false positive rates across different operating systems, with slightly higher rates observed in Linux environments. Resource utilization assessments indicated moderate CPU and memory usage during peak detection activities, suggesting suitability for real-time deployment. However, a gradual decline in detection rates for successive generations of polymorphic ransomware highlights the need for continuous refinement. The GCPDA framework represents a significant advancement in ransomware detection, offering a robust and adaptable solution to the evolving threat landscape.

This paper investigates the motivations for developing underwater gait assistance robots and explores methods that leverage the benefits of partial submersion for rehabilitation. An overview of the current state of underwater gait rehabilitation robots is provided, highlighting the successes and limitations of existing designs. The study surveys key approaches in gait characterization, actuation, sensing, system modeling, and control strategies to identify those best suited for underwater environments. Emphasis is placed on how buoyancy and drag can be harnessed during hydrotherapy to improve gait rehabilitation outcomes. The review identifies research gaps and outlines future opportunities for designing effective underwater robotic systems that complement therapeutic needs.

The design, fabrication, and demonstration of a novel Silicon Carbide Low Gain Avalanche Detector (4H-SiC LGAD), exhibiting an ultra-fast time response and excellent time resolution, are reported. The use of field plates is proposed to suppress the high electric field caused by the negative bevel-etched angle in 4H-SiC LGADs based on TCAD simulations. Experimental measurements confirm that the field plate significantly increases the breakdown voltage of the 4H-SiC LGADs. Gain and time resolution are measured by using the ultraviolet transient current technique (UV-TCT) showing that 4H-SiC LGADs possess excellent timing performance, with a time resolution better than 35 ps for the injected laser signal with single minimum ionizing particle (MIP) charges generation at room temperature. Additionally, the gain suppression effect in a 4H-SiC LGAD is observed for the first time.

The Indian military healthcare system plays a crucial role in safeguarding the health and well-being of those who serve in the military, including active personnel, retired veterans, and their dependents. This extensive network of hospitals and medical facilities spans across all branches of the armed forces—Army, Navy, and Air Force—providing specialized medical care. Despite its broad reach, the system faces challenges, particularly within Outpatient Departments (OPDs), which handle a substantial volume of daily visits. This workload tends to spike during health crises, such as pandemics. Managing patient flow effectively remains a persistent challenge for military hospitals. Traditional queuing strategies, such as the first-come, first-served and token-based systems, fall short of handling the complexity of patient prioritization, doctor efficiency, and real-time workload adjustments. These methods often lead to inefficiencies like extended waiting periods, uneven workload distribution among medical personnel, and suboptimal resource use. The manual system’s inability to adapt to changing patient demands and the urgency of medical needs further exacerbates these issues. There is an urgent need for improved queue management solutions in military hospital OPDs. A dynamic queuing algorithm that can adapt to real-time data—such as patient arrivals, doctor availability, and consultation times—could significantly enhance patient flow, reduce waiting times, and improve overall system efficiency. By implementing a flexible queuing approach, military hospitals can better address the needs of their beneficiaries, thereby enhancing both operational outcomes and patient satisfaction.

The Indian military healthcare system is tasked with ensuring the physical and mental well-being of service members, veterans, and their dependents. Despite its extensive network of specialized facilities, the system faces persistent challenges in managing patient flow, particularly in the Outpatient Departments (OPD). Current queuing methods, predominantly first-come, first-served and token-based systems, fail to address the complexities of patient prioritization, real-time doctor efficiency, and dynamic workload balancing, resulting in prolonged waiting times and inefficiencies.This research introduces a novel dynamic queuing algorithm designed to optimize patient flow in military OPDs. The algorithm combines Weighted Round Robin (WRR) for priority-based assignment, a Tribonacci-based aging mechanism to dynamically elevate patient priority with waiting time, and real-time load balancing to ensure equitable resource utilization. The study evaluates the algorithm’s performance through a simulation with heterogeneous doctor efficiency and diverse patient priorities. Results demonstrate significant reductions in patient waiting times and improved workload distribution among medical personnel, showcasing the algorithm’s potential to enhance operational efficiency and patient satisfaction in military healthcare facilities.By addressing limitations in traditional systems and leveraging a multi-dimensional optimization approach, this study contributes a scalable, adaptive framework for efficient queuing in high-demand healthcare environments. This innovation not only supports the operational objectives of military hospitals but also sets a precedent for broader applications in complex healthcare systems.

Modern large-scale computing systems improve job completion and energy efficiency by executing tasks in parallel. However, the Long Tail problem where a few slow-performing tasks, or stragglers, delay overall execution remains a key challenge in cloud computing. This paper introduces an Energy-aware Straggler Task Management Framework for sustainable cloud environments. The framework integrates predictive analytics with an evolutionary Neural Network to identify stragglers based on historical resource usage. A self-adaptive differential evolution algorithm is employed to optimize the neural network weights, enhancing prediction accuracy. Tasks are classified as stragglers or non-stragglers, with stragglers further divided into Resource Hunter and Long-Tail categories. A resource management policy dynamically assigns resources based on task type to enhance parallelism and sustainability. Extensive simulations were performed using the Google Cluster Dataset to analyze the effectiveness of the EaSTM framework. The obtained results highlight the framework's potential for improving efficiency and sustainability in cloud environments.