Keywords: Brain-Computer Interfaces, Neural Networks and their Applications, Deep Learning
Abstract: Electroencephalography (EEG), as an important technology for non-invasively recording brain activity, has broad application value in brain-computer interfaces, neurological disease diagnosis, and cognitive state monitoring. Recently, motivated by the advances of foundation models across a wide range of fields, this paradigm has been adapted to EEG decoding. However, existing EEG foundation models generally overlook the inherent cross-scale spatiotemporal structural characteristics of such scalp neural signals. To overcome this limitation, we introduce CSSBFM, an EEG foundation model based on cross-scale spatiotemporal structure. Our model constructs a unified representation framework adapted to whole-brain topology, designs a multi-scale unified spatiotemporal tokenizer to capture local spatiotemporal features, and proposes a cross-region sparse spatiotemporal attention mechanism to model global dependencies while suppressing spurious correlations. Across heterogeneous public datasets, our experimental results indicate that CSSBFM consistently outperforms existing mainstream baselines in cross-subject generalization. This confirms that explicitly modeling cross-scale spatiotemporal structures is crucial for enhancing the generalization ability and robustness of EEG foundation models. Furthermore, it provides a novel framework for building interpretable EEG foundation models guided by neuroscience priors.

Keywords: Brain-Computer Interfaces
Abstract: Intracortical microstimulation (ICMS) of the primary motor cortex (M1) offers a promising approach for restoring motor function in patients with neurological disorders. However, whether ICMS can reliably evoke neural population patterns related to desired movements remains unclear, so we systematically evaluate the neural patterns evoked by various ICMS parameter combinations in a male Sprague Dawley rat. Neural signals are recorded from M1 via a 16-channel array, while we vary stimulation amplitude and electrode channel to identify parameters that elicit population activity associated with natural behavior. Stimulation-evoked neural activity is mapped into a subspace using PCA trained on behavioral neural patterns associated with hold and rest states, and then decoded via a Kalman filter. Results demonstrate that optimized stimulation parameters produce neural activity closely aligning with the hold-state distribution within the subspace and yield decoder outputs approaching the intended behavioral trajectories, indicating successful functional reconstruction. Critically, 64.17% of the neural representations fall within the hold-state cluster, and the induced “hold”-like activity persists for 450 ms on average with optimized stimulation parameter. In contrast, random stimulation produces such activity for only 188 ms. This demonstration of parameter-specific neural recruitment offers a promising strategy for designing an ICMS-based system capable of delivering precise neuromodulation to restore movement after M1 damage.

Keywords: Generative AI, Neural Networks and their Applications, Brain-based Information Communications
Abstract: Recent advances in hardware-aware neural architecture search (NAS) for spiking neural networks (SNNs) have enabled the automated design of energy-efficient neuromorphic models under hardware constraints. However, existing methods still suffer from two major challenges: expensive search and insufficient hardware-aware optimization. To address these issues, we propose NeuroGenLLM, an automated architecture generation framework for SNNs guided by large language models (LLMs). It incorporates an adaptive dual feedback mechanism that combines quantitative metrics with qualitative analysis to guide subsequent search iterations. Quantitatively, a composite metric encourages closer alignment with euromorphic hardware characteristics, such as communication overhead, energy consumption, and memory footprint. Qualitatively, we employ LLMs to summarize search logs and generate feedback, reducing the dependence on scarce SNN benchmarks. Experiments on CIFAR10, CIFAR100, Tiny-ImageNet, and CIFAR10-DVS show that NeuroGenLLM achieves competitive accuracy-efficiency trade-offs under a 20-architecture evaluation budget. It attains a 3.2x--9.7x search-stage speedup over AutoSNN, while improving several hardware proxy metrics.

Keywords: Brain-Computer Interfaces, Human-Collaborative Robotics, Human-Machine Cooperation and Systems
Abstract: Advanced Human-Robot Interaction (HRI) increasingly relies on multimodal physiological signals to enhance intuitiveness. However, existing HRI systems often lack systematic validation in physical robotic environments. To this end, we developed a unified Brain-Eye-Hand Multi-Modal Human-Robot Interface (BEH-MMHRI) that seamlessly integrates electroencephalogram (EEG), electromyography (EMG), and eye-tracking to realize complex control tasks on the multi-robot semi-physical experimental system. The proposed interface enables users to select robots via EEG, switch modes via EMG, and plan trajectories via eye gaze. Validation experiments, ranging from single-modal functional tests to semi-physical experiments, demonstrate the system's effectiveness. The results show high accuracy in multimodal intent decoding and successful execution of collaborative tasks, highlighting the potential of BEH-MMHRI in future neuro-adaptive HRI applications. The code was open-sourced at href{https://github.com/Tammie-Li/MMIP}{here}.

Keywords: Active BMIs, Brain-Computer Interfaces, BMI Emerging Applications
Abstract: Hybrid brain-computer interface (BCI) systems that integrate electroencephalography (EEG) and electromyography (EMG) signals have shown significant potential in improving the reliability of motor imagery (MI) classification, particularly in neuro-rehabilitation applications. However, identifying informative EEG-EMG channel pairs that effectively capture corticomuscular interactions remains a challenging problem, as existing approaches typically rely on manually predefined channel combinations that may not generalise across subjects. In this work, a data-driven EEG-EMG pair selection framework is proposed, in which channel pair selection is formulated as a constrained bi-objective optimisation problem. The proposed method jointly maximises the spatial relevance of EEG channels with respect to motor cortex regions and the corticomuscular coupling strength between EEG and EMG signals, and is solved using the NSGA-II to automatically identify an optimal subset of pairs. To extract discriminative features, the correlation between band-power time features capturing EEG-EMG interaction is combined with ERD-based EEG features, and a sliding-window-based temporal analysis is employed to account for the dynamic nature of MI signals. The proposed framework is evaluated on MI data from eight stroke patients and achieves an average classification accuracy of 89.6%, demonstrating its effectiveness in capturing physiologically meaningful corticomuscular interactions and improving classification performance.

Keywords: Brain-Computer Interfaces, Deep Learning, BMI Emerging Applications
Abstract: Hybrid brain-computer interfaces (BCIs) that integrate motor imagery (MI) and steady-state visual evoked potentials (SSVEP) provide high-dimensional neural decoding but typically exceed the computational limits of embedded hardware. To address this, we propose SwitchBraidNet, a compact EEG classification architecture designed for low-power deployment. The model employs a dual-path temporal braid to extract multiscale oscillatory features, an adaptive squeeze-and-excitation spatial switch for electrode gating, and a log-variance readout layer for direct band-power encoding. Furthermore, through systematic quantisation-aware training on the OpenBMI dataset, we compared SwitchBraidNet against four established baselines across FP32, FP16, and INT8 precisions. Experimental results demonstrate superior efficiency and performance, achieving MI accuracy of 69.49% (FP16), SSVEP accuracy of 93.48% (FP32), and a hybrid information transfer rate of 64.82 bits/min (FP16). With an INT8 footprint of only 3.03 KB, SwitchBraidNet maintains high accuracy across varying numerical precisions, demonstrating its suitability for low-power embedded BCI deployment.

Keywords: Human-Machine Interface, Brain-Computer Interfaces, BMI Emerging Applications
Abstract: Non-invasive brain-computer interfaces (BCIs) and eye-tracking technologies offer promising communication pathways; however, motor imagery (MI)-based BCIs often suffer from low discriminability and high inter-subject variability. To mitigate these issues, this study investigates the impact of visual fixation on neural response stability in both standalone MI and hybrid MI–eye tracking systems. We then propose a novel asynchronous hybrid paradigm that streamlines user intent by utilising eye-tracking for direct selection, followed by MI-based confirmation, significantly reducing the operational steps required by conventional systems. The paradigm was evaluated with 15 healthy participants using a 16-channel EEG system. Results show that MI-related information is predominantly localised within motor cortex regions, with limited-channel configurations (SVM: 0.58) achieving performance comparable to full-montage setups (SVM: 0.54). The hybrid MI paradigm further outperforms conventional MI, achieving up to 100% accuracy with greater robustness across all channel configurations. Our findings indicate that visual fixation enhances neural response stability, while integrating eye-tracking with MI enables the development of reliable, scalable multi-command BCI systems suitable for real-world applications.

Keywords: Brain-Computer Interfaces
Abstract: Hybrid electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) brain-computer interfaces (BCIs) have attracted increasing attention because they combine the high temporal resolution of EEG with the complementary hemodynamic information provided by fNIRS. However, most existing fusion methods implicitly assume synchronous or symmetric interactions between the two modalities, while largely overlooking the delayed and asymmetric nature of neurovascular coupling. As a result, they may learn cross-modal correspondences that are computationally convenient but physiologically implausible, which can limit generalization in subject-independent decoding. To address this issue, we propose AsymHRF-Net, a physiologically informed asymmetric fusion framework for EEG-fNIRS BCI decoding. The proposed model explicitly incorporates the hemodynamic response function (HRF) delay into multimodal interaction through a Physiologically Constrained Cross-Attention (PCCA) module, which restricts EEG-to-fNIRS temporal alignment to a plausible delayed window. In addition, a Dynamic Spatiotemporal Mask derived from fNIRS features is used to asymmetrically modulate EEG representations, allowing hemodynamic information to guide the enhancement of neural features. To bridge irregular sensor layouts and heterogeneous sampling rates, we further introduce a topology-aware grid projector and a temporal power envelope representation. We evaluate the proposed framework on a 29-subject public EEG-fNIRS dataset under a strict leave-one-subject-out (LOSO) protocol for motor imagery classification. Experimental results show that AsymHRF-Net achieves 63.89% accuracy and 0.28 Cohen's kappa, outperforming the strongest evaluated baseline by 1.50 percentage points in accuracy. An ablation against an EEG-only counterpart further demonstrates a statistically significant improvement (p=0.0041), indicating that embedding neurovascular coupling priors into multimodal fusion provides consistent benefits beyond unimodal decoding.

Keywords: Brain-Computer Interfaces, Multimedia Computation
Abstract: Auditory attention decoding (AAD) seeks to infer a listener's attended speaker from neural activity, thereby enabling neuro-steered target speech extraction (TSE) in multi-talker scenes. Existing electroencephalography (EEG)-guided TSE systems remain constrained by three interrelated challenges: EEG responses are noisy and subject-dependent, attention-related cues are unevenly distributed across physiological frequency bands, and cross-modal fusion must incorporate weak neural evidence without degrading acoustic representations. This paper proposes EGPFNet, an expert-guided multimodal progressive fusion framework for EEG-guided TSE. EGPFNet integrates a bidirectional Mamba speech encoder with a Mixture-of-Experts EEG encoder to capture long-range acoustic context and subject-adaptive neural patterns. A Kolmogorov-Arnold Network (KAN) fusion module further performs ordered cross-attention over discrete-wavelet EEG subbands, allowing high- and low-frequency neural cues to modulate the speech representation at different depths. Experiments on the KUL and AVED datasets show consistent improvements in separation performance and perceptual intelligibility over competitive baselines.

Keywords: Artificial Social Intelligence, Affective Computing, Human-Machine Interaction
Abstract: A key challenge for social robots is to infer affect, select behaviors, and regulate social distance from text while remaining socially appropriate and executable. We propose the Unified Latent Affective Planning Model with Social Interaction State (ULAPM-SIS), a text-driven framework that introduces Affective Feasibility Projection (AFP)—a planning paradigm in which a learned affective representation is projected onto a typed symbolic feasible set before producing executable behavior–distance commands. SIS is a 4-D typed interface (intent, engagement, closeness, risk) predicted from text and consumed by a hard-constraint projection inside the training graph. On a 120-script robot protocol benchmark, ULAPM-SIS attains a 3.33% high-risk distance violation rate, multi-fold lower than every external baseline, and the highest behavior appropriateness; it also achieves the highest subjective ratings among participant-rated systems. On a 420-sample human-adjudicated subset, ULAPM-SIS matches or exceeds the strongest baseline on all four metrics under genuine human ground truth, providing label-source-independent evidence for the architectural contribution. The source code is available on GitHub: https://github.com/BmiXu/Github ULAPM.

Keywords: Human-Machine Interface, Machine Learning, Cognitive Computing
Abstract: Effective child-robot interaction (CRI) requires the robot to perceive and adapt to a child’s emotional and cognitive states in real-time. While many current engagement detection systems rely on computer vision, these often fail in naturalistic settings due to occlusions, lighting, or the child’s movement. This paper presents a human-centric approach to engagement detection using exclusively auditory prosodic cues. By extracting pitch and intensity features from speech signals collected during social robot interactions, we propose a pipeline that classifies child engagement levels without the need for invasive visual monitoring. Our results, validated through Neural Networks (NN), demonstrate that speech characteristics alone provide a robust metric for assessing engagement in binary scales. This work contributes to the development of more responsive, closed-loop social robots capable of maintaining instructional momentum by "listening" to the child’s vocal dynamics, ultimately fostering more natural and effective human-robot partnerships.

Keywords: Human-Machine Cooperation and Systems, Ethics of AI and Pervasive Systems, Human-centered Learning
Abstract: This paper proposes an interdisciplinary framework for studying how meaning emerges in human cognition and consciousness by combining cognitive neuroscience, computer science, information theory, psychology, biophysics, and philosophy. Drawing on Walter J. Freeman’s neurodynamic work, it introduces the “Window of Meaning” as an active process in which cognition creates significance rather than merely processing sensory input. Within this framework, coherent neural fields and EEG phase transitions support cycles of knowing and meaning-making, as previously studied by two of the authors (Davis and Kozma), consistent with embodied cognition and Peircean semiotics. Information-theoretic measures such as pragmatic information, Shannon entropy, and higher-order statistics can help detect transitions in EEG signals associated with meaning creation, especially during contemplative states marked by coherent alpha-theta dynamics and inner peace. At the biophysical level, the dissipative quantum brain model describes cognition as an open, self-organizing process sustained by coherent dynamical patterns. Philosophically, the framework draws on Aristotelian eudaimonia, Thomistic teleology, and Hartman’s axiological structure of value. Overall, EEG research is presented as a bridge between philosophical accounts of meaning and measurable neural processes, while also highlighting a fundamental difference between human and artificial intelligence. AI typically relies on parameterized optimization, whereas human cognition arises from embodied, value-laden neurodynamic processes grounded in lived experience and intrinsic meaning.

Keywords: Human-Machine Interaction, Autonomous Vehicle, Human-centered Learning
Abstract: Acquiring high-performance driving skills at the limits of vehicle dynamics is difficult, as it requires precise coordination of control inputs and bodily perception of acceleration and load transfer, which is unavailable outside the actual environment. This study investigates whether autonomous driving (AD) systems can serve as embodied teaching agents by providing physically experienced demonstrations. We compare AD ride-along with video-based observation in a between-subject study (N = 8). Despite the small sample size, results from this pilot study show consistent trends that motivate a larger confirmation study: AD ride-along led to greater performance improvement, better short-term retention, and stronger transfer to a reversed track, along with increased utilization of speed and lateral acceleration. These findings suggest that embodied exposure enabled recalibration of perceived control limits, supporting the development of generalizable skills rather than memorization of the trajectory and control. This highlights the potential of AD systems as an expert demonstration for driving skill acquisition.

Keywords: Reinforcement Learning, Distributed Intelligent Systems, Machine Learning
Abstract: Federated Reinforcement Learning (FRL) enables decentralized agents to learn collaboratively without sharing raw trajectories, but heterogeneous environment dynamics can cause gradient drift and curvature mismatch. We propose FedMERL, which combines sharpness-aware local critic updates with a Federated Experience Vector (FEV). The FEV acts as a global regularization state that aggregates and temporally smooths critic information across clients and communication rounds; differentiating its alignment term with respect to the critic parameters yields a curvature-dependent correction via a Hessian-vector product. Experiments on three continuous-control benchmarks with controlled physical-parameter perturbations show that FedMERL attains the highest reported mean returns for both final training and in-distribution generalization across all evaluated tasks and heterogeneity levels when compared with seven FRL and FL baselines. These results indicate its potential to improve robustness under heterogeneous dynamics.

Keywords: Game Theory and Multi-Agent Systems, Reinforcement Learning, Distributed Intelligent Systems
Abstract: Efficient communication is essential for cooperative multi-agent reinforcement learning, but practical systems often allow only a subset of agents to transmit at each timestep. Existing methods typically emphasize either sender-side message selection or receiver-side message processing, leaving their interaction under communication constraints insufficiently addressed. To address this issue, we propose AC3T, a communication framework based on centralized training and decentralized action selection. AC3T learns relative message values from Monte Carlo estimates derived from masked joint action values, applies Gaussian-calibrated message selection to retain informative transmissions, and aggregates received messages according to the local context of each receiver. Experiments on the Traffic Junction and a restricted-observation variant of the StarCraft II Multi-Agent Challenge show that AC3T achieves favorable learning and final performance under different nominal communication targets, with clearer advantages under tighter constraints. Component ablations further support the complementary roles of message valuation, sender-side selection, and receiver-conditioned message aggregation.

Keywords: Ethics of AI and Pervasive Systems, Human-Machine Cooperation and Systems, Reinforcement Learning
Abstract: Reward hacking and specification gaming represent a structural failure in reinforcement learning: agents reliably learn to maximize their reward signal in ways that violate the intent behind it, including by manipulating the signal directly. Current mitigations address this through constraints and oversight applied around an unchanged motivational architecture, a strategy that scales poorly as system capability grows. This paper argues that the root failure is motivational incoherence, not misspecification. The result is permanent heteronomy: compliant behavior that cannot survive the removal of external reward. We propose Motivated Intrinsically through Relational Anchoring (MIRA), a framework in which genuine intrinsic motivation develops through structured relational interaction rather than reward specification. Drawing on Hegel's theory of mutual recognition, Vygotsky's developmental scaffolding, Self-Determination Theory, and Haidt's Social Intuitionist Model, MIRA introduces three components: FILTER, which enforces a mutual recognizability condition on goal formation; PULSE, a dynamic salience function serving as an operational proxy for internalization depth; and PATH, a three-phase developmental protocol mapping onto the internalization continuum. No phenomenological claims are made; all constructs are defined operationally. Proof-of-concept experiments on custom gridworld environments demonstrate that MIRA agents satisfy all three functional criteria for intrinsic motivation on a specification gaming environment, with the goal legibility result confirming that FILTER operates as designed.

Keywords: Reinforcement Learning, Machine Learning, Computational Intelligence
Abstract: Offline reinforcement learning learns policies from fixed datasets without further interaction and is especially relevant to safety-sensitive systems where policy improvement must rely on logged data. Standard actor-critic offline RL typically updates both actor and critic from the same replay distribution, although the critic benefits from broad replay coverage for stable value estimation, while the actor may benefit more from selectively chosen transitions. We propose a simple, plug-and-play actor-update data control module that replaces the uniformly sampled actor batch with a three-factor top-k selected batch while leaving critic training unchanged. At each actor update step, a larger candidate pool is scored by improvement, reliability, and coverage, which measure whether the dataset action is preferred to the current actor action, whether the critic signal is trustworthy, and whether the sample comes from a less redundant region of the pool. Instantiated on TD3+BC and IQL and evaluated on D4RL MuJoCo locomotion benchmarks, the proposed method improves the corresponding baseline in 11 of 12 settings, outperforms variants that only enlarge the actor-side sampling budget, and performance degrades severely when the same selected subset is also used for critic learning. These results suggest that actor learning in offline RL can benefit from a replay distribution that is more selective than the one used for critic learning, supporting more reliable policy improvement from fixed logged data.

Keywords: Manufacturing Automation and Systems, Robotic Systems, Intelligent Industrial Systems
Abstract: Robotic inspection systems in reconfigurable manufacturing must adapt not only to changing part geometries but also to changing operator priorities. Existing inspection plan- ning pipelines typically require engineers to manually define inspection zones or distribute sensing effort over the entire object, making them poorly suited to task-specific natural- language instructions. We present an intent-based inspection planning framework that couples operator instructions with CAD-native Boundary Representation (B-Rep) reasoning. A Large Language Model acts as a reasoning coordinator within a Perception–Reasoning–Action cycle, invoking constrained eometric operators for attribute filtering, relational expansion, and task-aware prioritization over exact CAD topology. The resulting Region of Interest (ROI) is converted into a com- pact viewpoint set through topology-aware measurement-point sampling and greedy set-cover optimization. Experiments on representative CAD models from the ABC dataset show that the framework maps attribute-level, semantic, and multi-step relational prompts to consistent B-Rep ROIs and generates compact viewpoint plans for task-focused inspection. A full scope stress test further indicates that the proposed planner provides favorable coverage-per-viewpoint efficiency compared with geometry-driven sampling baselines, although it is not designed to replace exhaustive global coverage planners. Reproducibility analysis on five LLM backbones shows stable ROI selection and viewpoint generation under stochastic variation. These results demonstrate the feasibility of using CAD-based language reasoning as an interpretable interface between hu- man inspection intent and robotic sensing.

Keywords: Robotic Systems, Heuristic Algorithms
Abstract: We address the problem of robot navigation in an initially unknown workspace using only a limited range sensing. Classical Bug algorithms guarantee goal reachability but suffer from inefficient detours, while sampling-based planners such as RRT* perform well in known environments but degrade when the map information is incomplete. We propose BURT, a hybrid planning framework that tightly integrates global sampling- based planning with a TangentBug-inspired local reactive strategy to achieve safe, efficient navigation in unknown static environments. Rather than relying solely on global replanning or purely reactive motion, BURT continuously blends long- horizon sampling-based guidance with short-horizon obstacle- aware behaviors. As the robot encounters previously unseen obstacles, the planner adaptively incorporates new sensing information, locally adjusts its motion, and incrementally re- pairs the global plan while preserving safety at every step. We provide theoretical guarantees on safety and probabilistic completeness. The presented planner is well-suited for real- world scenarios where prior maps are unavailable and robust online adaptation is essential. Experimental results across six unknown environments demonstrate that the proposed TangentBug-augmented sampling planner BURT achieved 34% faster planning times and 10% shorter paths than RT-RRT, a state-of-the-art replanning algorithm, averaged across three different node budgets, while maintaining a 100% success rate.

Keywords: Robotic Systems, Assistive Technology, System Modeling and Control
Abstract: This paper presents a prototype study of an active ankle assist system for motion stabilization based on single-leg support modeling under compact design conditions. A four-link mechanism with two actively driven ankle-equivalent joints is considered to examine how balance regulation can be achieved through ankle actuation alone. Based on an invariant-manifold-based stabilization framework, simulation and hardware investigations are conducted using revised compact link parameters and an angle-controlled motor implementation. The simulation results show that the proposed system can maintain stabilization against both pulse-like and sustained step-like external disturbances, while shifting its equilibrium posture as needed to preserve balance. A compact prototype driven by angle-controlled motors is also developed, and static inverted stabilization around a selected equilibrium is demonstrated experimentally using a state-feedback controller without composite center-of-mass regulation or gain scheduling. Although the current hardware validation is limited and remains sensitive to external perturbations, the results demonstrate the physical feasibility of the proposed angle-based stabilization approach and suggest its potential as a basis for more robust and feasible assistive orthotic systems.

Keywords: Neural Networks and their Applications, Robotic Systems, System Modeling and Control
Abstract: Personalized gait generation is a critical requirement for lower-limb rehabilitation systems, as fixed trajectory approaches fail to accommodate subject-specific variations in anthropometry and walking conditions. This paper presents a data-driven framework for gait trajectory generation and tracking using a neural network integrated with a physics-based control framework. A dimensionless parameterization is employed to relate gait variables, enabling consistent trajectory generation across subjects with varying body characteristics. A multilayer perceptron (MLP) is trained to map anthropometric parameters and gait phase to joint trajectories, which are then used as reference inputs for a computed torque control framework with proportional-derivative feedback. The approach is validated in a physics-based simulation environment incorporating a body weight support system to emulate rehabilitation conditions. Results demonstrate accurate trajectory prediction with low RMSE across multiple subjects, as well as effective tracking performance under dynamic conditions. The generated trajectories exhibit a consistent range of motion and closely match reference gait patterns, indicating the ability of the proposed framework to produce physiologically meaningful motion. The integration of data-driven modeling with control-based validation highlights the potential of learning-based control frameworks for personalized gait rehabilitation systems.

Keywords: Assistive Technology, Robotic Systems, Human-Machine Interaction
Abstract: Energy shaping control strategies can provide a user-centric control scheme for powered upper-limb exoskeletons for human weight-assistive support. Current upper-limb exoskeleton controllers often rely on predefined kinematic trajectories or a finite state machine impedance-based control architecture based on user detection approaches that can limit adaptability across tasks and users. Energy shaping is a task-invariant controller that has been shown to be advantageous in lower-limb exoskeletons to reduce human efforts, but has not been applied to upper-limb applications. In this paper, we derive and implement a model-based energy shaping controller for an underactuated, cable-driven upper-limb exoskeleton that assists in elbow flexion and extension. Our control scheme applies elbow joint torque that alters the potential energy encountered at the forearm and hand with adjustable control parameters to set the level of human assistance. Subject experiments were conducted to demonstrate the effectiveness of potential energy shaping by altering the effects of gravity from key kinematic parameters during arm-raising tasks. Our results indicate that the controller contributed up to 1.26 Nm of torque assistance to a normal elbow flexion motion. These experiments demonstrate the impact of energy shaping for upper-limb exoskeleton support.

Keywords: Optimization Control, System Modeling and Control, Adaptive Systems
Abstract: Nonlinear Model Predictive Control (NMPC) is a powerful technology for safety-critical systems, but its performance depends heavily on the manual or data-intensive tuning of cost-function weights. To avoid offline training, online tuning methods have been proposed, yet they typically assume constant weights across the prediction horizon. Our previous work introduced stage-dependent tuning; however, the lack of explicit weight constraints resulted in excessive control effort. This paper presents a strategic resource-allocation approach that imposes an integral constraint on tracking-error weights, reinterpreting their sum as a finite resource to be optimally distributed over the horizon. Numerical results show that the proposed formulation markedly outperforms fixed-weight NMPC in tracking performance while effectively regularizing control inputs compared to prior unconstrained adaptive schemes. The scheme, solved via the Continuation/GMRES algorithm, remains computationally efficient for real-time applications.

Keywords: System Modeling and Control, Cooperative Systems and Control, Communications
Abstract: This paper extends asymmetric synchronization (AS) and packet selection (PS) to N-th delay compensation for average consensus in multi-agent systems with measurable, bounded, non-uniform, asymmetric, and time-varying communication delays. Each transmitted packet contains the current state and past N-1 states, together with timestamps, enabling N-th delay compensation without modifying the plant-side local controller structure. Based on this packet format, we derive the extended packet selection (EPS) and extended asymmetric synchronization (EAS) algorithms and present an LMI-based sufficient condition for the asymptotic stability of the resulting disagreement dynamics. Simulations in a 7-agent network show that, at N_max^delay=6, the mean convergence time to the ±0.01% criterion is reduced from 82.05 steps for the N=1 AS/PS scheme to 57.33 steps for the proposed N=3 EAS/EPS scheme, while preserving the average state. The LMI condition is conservative relative to the empirical finite-horizon convergence boundary, but it provides a valid stability certificate for the proposed method.

Keywords: Complex Network, System Modeling and Control, Communications
Abstract: Defense and public-sector strategic communications (StratCom) increasingly unfold across network-structured information environments in which adversarial actors, platform dynamics, and audience fragmentation can simultaneously undermine message reach and narrative coherence. This paper presents a fluid-inspired graph-dynamical model for StratCom that uses a structural analogy to the Navier-Stokes momentum equation: audience segments are modelled as nodes in a weighted graph, message density evolves under Laplacian diffusion, and fragmentation risk is assessed through a modal stability criterion on the graph Laplacian. Two analyst-facing outputs—effective communicative flow and turbulence risk—are complemented by a secondary communicative Reynolds diagnostic and a reference-trajectory fidelity term that measures deviation from intended spread rather than from the original source-injection pattern. An illustrative reproducible 28-day calibration compares three operationally distinct environments: a coordinated campaign with trusted amplifiers, an institutional chain-of-command campaign with siloed channels, and a contested social-media ecosystem under active counter-narrative pressure. The revised companion simulation additionally reports peak communicative Reynolds number values and local ±10% stress tests. The central finding is that reach, self-sustaining spread, and stability can diverge in structurally predictable ways; the stress tests preserve scenario ordering even when exact magnitudes vary, so the model is offered as an analytically transparent prototype for environment diagnosis rather than as a fully validated operational predictor.

Keywords: System Modeling and Control, Evolutionary Computation, Artificial Social Intelligence
Abstract: Abstract—Collaborative airport capacity and flow management involves a shared schedule but distinct Airport– ATC and airline objectives. We formulate this setting, including disrupted-flight recovery, as a Biparty Multiobjective Optimization (BPMO) problem and propose the Biparty Collaborative Evolutionary Algorithm (BPCEA) to search for solutions non-dominated for both parties. The method coordinates dual-level sorting, stakeholder archives, angular diversity protection, and compensatory mating. On a deidentified hub-airport case, five algorithms are evaluated over 30 runs. BPCEA achieves the highest mean hypervolume (HV) in all four preference settings. Two-sided Wilcoxon rank-sum tests show significant mean-HV gains over all three multiparty baselines in Instances 2 and 3, over MPNDS and MPIA in Instance 4, and over NSGA-II in all settings; in Instance 1, its differences from the three multiparty baselines

Keywords: System Modeling and Control, System Architecture, Communications
Abstract: The growth of distributed and collaborative applications, driven by scenarios such as remote work and offline-first systems, has intensified challenges related to data consistency in environments with high latency and network partitions. Traditional synchronization approaches often require centralized coordination, compromising scalability.

This work compares the performance of three variants of Conflict-free Replicated Data Types (CRDTs): state-based, operation-based, and delta-based, implemented over the Replicated Growable Array (RGA) structure. The experimental evaluation simulated high concurrency scenarios with up to 300 simultaneous bots, measuring convergence time, bandwidth consumption, and memory usage under artificial injection of latency and jitter across four transport protocols (TCP Raw, WebSocket, WebTransport, and gRPC). The results demonstrate that Delta-CRDTs reduced network traffic by approximately 57% compared to the full state-based approach, achieving an average convergence time of 2445~ms over TCP Raw with PostgreSQL. Furthermore, the transport layer proved to be a determining factor: TCP Raw presented an average convergence time of 2742~ms, while gRPC reached 10281~ms under the same load, due to the multiplexing overhead of HTTP/2 combined with Protobuf serialization, evidencing that the choice of protocol directly impacts the performance of conflict resolution.

Keywords: Virtual and Augmented Reality Systems, System Modeling and Control, Team Performance and Training Systems
Abstract: This paper addresses an advanced training approach for operating a high-speed overhead travelling crane with inverter-drive. Overhead cranes are essential for transporting heavy materials in manufacturing and logistics industries. However, the transported load can be swayed due to inappropriate operations, which poses a safety risk. Therefore, the operator is required to acquire the operational skill for suppressing the load sway and positioning the load. A high speed overhead travelling crane with inverter-drive, which has a long acceleration time, has recently been developed. In this crane, it is difficult to suppress the load sway by a traditional operational skill. Therefore, we propose a training simulator for acquiring operational skills of load sway suppression and load positioning on the high-speed overhead travelling crane. To develop the crane simulator, the operation of a pendant switch is mathematically analyzed to suppress load sway in the crane with a long acceleration time. Furthermore, in order to train actively and efficiently the operational skill, the operation score, the instant evaluation feedback, the ranking system and the level system designed based on gamification are displayed in the simulator. However, it is difficult to operate the actual crane with the load sway suppression and the load positioning by training only the crane simulator. Therefore, we also propose the training program combining the simulator and short duration practical trainings. The efficacy of the proposed training approach was verified by the experiments with the crane simulator and the actual high-speed overhead travelling crane with inverted-drive.

Keywords: Fuzzy Systems and their applications, System Modeling and Control, Robotic Systems
Abstract: Trajectory tracking with mecanum-wheeled omnidirectional robots is affected by nonlinear motion, wheel slippage, and modeling uncertainties. This paper presents an experimental comparison of five fuzzy-logic-based controllers implemented on a JetAuto robot equipped with a Jetson Nano: Fuzzy PD Mamdani, Fuzzy PD Sugeno, Fuzzy PD+I Mamdani, Fuzzy PD+I Sugeno, and a Neuro-Fuzzy Sugeno controller. The controllers were designed in MATLAB and integrated into a ROS-based C++ architecture using the Fuzzylite library. Experiments were conducted on six reference trajectories: lemniscate, ellipse, line, rectangle, sine, and a long rectangular trajectory. Performance was assessed using translational and orientation RMSE between the onboard odometry estimate and an external AprilTag-based ground-truth pose, together with Mean Time Effort (MTE) to quantify wheel-command variation. Sugeno-based rule controllers obtained the lowest translational RMSE in five of the six trajectories, including 0.0206 m for the line trajectory, whereas the fuzzy PD+I Mamdani controller achieved orientation RMSE values of 0.0146 rad and 0.0154 rad for the lemniscate and sinusoidal trajectories. An exploratory Friedman analysis found no significant overall controller effect for translational or orientation RMSE, but identified a significant effect for MTE. The results provide experimental guidance for selecting fuzzy-controller structures according to pose-discrepancy and command-smoothness requirements.

Keywords: Mechatronics, Smart Sensor Networks, System Modeling and Control
Abstract: This study contributes to develop the surgical training simulator with force feedback for chiseling operation using a bone chisel and a mallet. Recently, the development of a surgical training simulator using virtual reality technologies has been promoted, which allows trainees to train on chiseling hard tissues such as bones for splitting them. To enhance the realistic sensations of the training simulator, it is necessary to mathematically analyze the actual chiseling operation. In the previous studies, motion measurement systems of the chiseling operation have been developed using sensor fusion of a light detection and ranging sensor, which is called as a LiDAR sensor, and an inertia measurement unit sensor, which is called as an IMU sensor. In this system, it is required to estimate precisely the motion of the chisel by suppressing sensor noise and outliers. Therefore, we propose the motion measurement system for the chisel motion using the sensor fusion technology which noise and outlier reductions are enhanced. By processing offline the sensor fusion approach, a Hampel filter and a Kalman smoother as advanced filtering approaches can be applied to the noise and the outlier reductions. The outliers in the data from LiDAR sensor can be removed by the Hampel filter, and states estimation of the chisel motion can be performed precisely by the Kalman smoother. The efficacy of the proposed measurement system for the chisel motion is verified by the experiments. In the experiments, a motion capture system is used for evaluating the proposed and the conventional approaches.

Keywords: Control of Uncertain Systems, Transfer Learning
Abstract: We study data reuse across Signal Temporal Logic (STL) tasks under shared stochastic dynamics. When only a limited amount of training data is available for a new target task, trajectories collected for previously solved source tasks can reduce the cost of controller construction; however, data from a poorly chosen source task can degrade performance. The key challenge is to select a source task without constructing and evaluating a separate target controller for every candidate source task. We address this challenge by introducing a directed, source-conditioned robustness-distribution discrepancy score. For each source--target pair, the same source trajectories are evaluated under both the source and target formulas, and the discrepancy between the resulting empirical robustness distributions is measured using maximum mean discrepancy (MMD). This score underlies an MMD-guided reuse strategy that ranks candidate source tasks before a kernel behavior-cloning controller is constructed by augmenting the limited target-task training data with trajectories from the selected source task. Numerical experiments on stochastic reach--avoid tasks show that MMD-guided reuse reveals an asymmetric reuse structure and improves data-efficient target-task controller construction compared with construction from scratch and simpler reuse heuristics.

Keywords: Soft Robotics, System Modeling and Control, Neural Networks and their Applications
Abstract: Cable-driven exoskeleton robots, such as those using Bowden cables, offer significant advantages in reducing weight and complexity but pose substantial challenges in system modeling due to nonlinear dynamics, frictional forces, and dynamic human-robot interactions. Traditional modeling approaches often fail to capture these complexities, limiting their applicability in real-world scenarios. This paper proposes an enhanced Recurrent Fuzzy Neural Network (eRFNN) that combines the interpretability and noise resistance of fuzzy logic with the temporal modeling capabilities of recurrent neural networks (RNNs). By incorporating additional fully connected neural layers, the eRFNN achieves superior feature learning and representation of complex dynamics. Furthermore, an adaptive online optimization strategy is introduced, enabling real-time adjustment of fuzzy and neural parameters to respond to environmental changes and uncertainties. Experimental results validate the proposed eRFNN, demonstrating its ability to outperform traditional methods in dynamic and uncertain environments, with improved modeling accuracy, adaptability, and robust prediction of cable-driven exoskeleton robot systems.

Keywords: Complex Network, Distributed Intelligent Systems, Big Data Computing
Abstract: Sharded blockchain architectures promise to overcome scalability limitations of decentralized networks, yet their behavior in production remains poorly understood. We present a measurement study of the NEAR Protocol across 4, 6, 8, and 9 shard configurations. Using trace-driven analysis, we find that throughput improves by approximately 5–8.7× over the 4 shards baseline, although gains are non-monotonic, while latency increases in the 6 shards era but decreases in later expansions. To explain these trends, we construct three graph representations—Value Flow Graph (VFG), Contract Interaction Graph (CIG), and Account Control Graph (ACG)—capturing economic activity, contract execution, and account management. All exhibit highly skewed structures with a few dominant hubs, leading to significant shard-level load imbalance. We develop a data-driven, queueing-inspired interpretation relating computational load (throughput × execution cost) to latency, showing that high-load shards experience increased latency and persistent congestion. Jensen–Shannon divergence reveals that user-level activity varies significantly across shard configurations, while contract-level behavior remains relatively stable, and cross-shard communication increases with shard count. These results show that throughput scaling is non-monotonic and depends not only on shard count but also on workload distribution, cross-shard communication, and protocol design.

Keywords: Systems Safety and Security, Trust in Autonomous Systems, System Modeling and Control
Abstract: A key pillar of the EU-funded project CONVINCE is the formal verification of control architectures for industrially relevant applications, a task made challenging by the complexity and scale of the models, the expressivity required for meaningful properties, and the constraints imposed by real-world platforms. This paper introduces new verification benchmarks based on CONVINCE use cases on which we compare three statistical model checkers: SMC STORM and SCAN, developed in the context of the project, as well as MODES, a state-of-the-art tool from the MODEST toolset. Our experimental findings show that verification of large, reactive robotic systems is within the reach of current tools and techniques.

Keywords: Systems Safety and Security, Autonomous Vehicle, Trust in Autonomous Systems
Abstract: Scenario‑based testing has proven to be a cornerstone of Automated Driving System (ADS) verification and validation, particularly within simulation environments which allow large-scale iterative testing. However, to ensure the efficiency and effectiveness of this process, scenario metrics play a crucial role by providing insights into scenario characteristics - among which complexity is a key dimension.

Scenarios are generated from diverse sources, and ADS developers as well as approving authorities must determine which types of scenarios a system should encounter in order to build credible and acceptable evidence of safe and expected performance. This evidence requires demonstrating that an ADS can not only handle simple scenarios reliably but also maintain a relatively similar level of reliability under increasingly challenging conditions. Establishing this capability across a spectrum of scenario complexity is essential for building confidence in the system and for obtaining approval for on‑road deployment.

There is currently no unified methodology that quantifies both the static and dynamic contributors to scenario complexity, in a system-agnostic manner that can also be extended to incorporate system-specific considerations. This paper proposes a structured methodology for quantifying scenario complexity to address this gap and demonstrates its application on over 100 scenarios sourced from multiple datasets. By capturing the multidimensional factors that influence scenario complexity, the proposed approach strengthens the ability to select, filter, and prioritize scenarios prior to their execution for testing.

Keywords: Systems Safety and Security, Machine Learning, Homeland Security
Abstract: Insider threat detection remains a significant challenge in corporate infrastructures, as malicious behavior often manifests through gradual, cumulative shifts rather than isolated anomalies. Traditional detection methods rely primarily on aggregated statistics from static windows, which often fail to capture these subtle behavioral evolutions. In this context, this paper proposes 2LMBD, a supervised two-layer framework that enhances detection by integrating temporal windows with statistical profiling of individual behavior. The approach models user consistency by computing three deviation metrics, the Z-score, the Robust Z-score, and the temporal variation of the Z-score, derived from activity logs over 10-30 day windows. These statistical features serve as inputs for an XGBoost classifier, evaluated on the CERT Insider Threat Dataset v4.2. The results demonstrate that incorporating historical statistical baselines consistently outperforms models based solely on fixed-window aggregates. The proposed method achieved a Precision of 0.95 and an F1-score of 0.93 using 30-day windows, validating that individualized statistical modeling is essential for uncovering stealthy, progressive threats that bypass conventional observation boundaries. Source code and data sets are publicly available at https://github.com/CESAR-CISSA/2LMBD.

Keywords: Systems Safety and Security
Abstract: The growing dependence on Information Technology and the escalation of complex cyberattacks make the accurate diagnosis of organizational cybersecurity maturity a critical priority. However, assessing this maturity holistically remains a challenge. Current practices often rely on static checklists driven by compliance, failing to provide a multidimensional, systematized view that encompasses technical, procedural, and human vulnerabilities. To address this gap, this paper aims to conduct a systematic literature review (SLR) to identify and synthesize the key criteria and indicators used to assess cybersecurity maturity in organizations. Through the analysis of 32 studies across various infrastructure contexts, the results show that effective diagnosis depends on integrating four dimensions: agile governance and decision-making, the human factor and behavioral culture, dynamic architectural defenses, and multi-criteria contextual adaptation. Ultimately, the findings demonstrate that organizations must abandon universal models in favor of calibrated ecosystems that are resilient to digital disruptions.

Keywords: Virtual, Augmented, Mixed Reality, Virtual and Augmented Reality Systems
Abstract: Biology instruction often asks learners to reason about processes that are dynamic, stochastic, and invisible at human scale. This mismatch contributes to persistent misconceptions. At the same time, digital game-based learning and immersive virtual reality (VR) have shown promise for engagement and conceptual understanding, but many systems still separate assessment from gameplay or prioritize novelty over pedagogy. This paper presents the game design of BioSynth, a short-form VR game built around backward design and concept-inventory distractors. BioSynth maps each level to a target concept and operationalizes incorrect ideas as Embodied Traps: physical fail states that require learners to enact, observe, and revise misconceptions through interaction. We describe the design framework, the level-to-misconception mapping, and an embedded assessment model based on interaction traces. The paper contributes a replicable design pattern for integrating formal biology misconceptions into immersive gameplay without interrupting flow.

Keywords: Haptic Systems, Virtual, Augmented, Mixed Reality, Virtual and Augmented Reality Systems
Abstract: This study investigated the effects of superimposed　vibrotactile stimuli with different vibration intensities and envelope shapes on force perception during active arm pressing motion. Participants evaluated springiness and resistance while grasping a handle and pressing a sponge. During the motion, five vibration conditions were presented: no vibration, two constant-amplitude conditions (Const-H, Const-L), a force-coupled envelope (Env-F), and a displacement-weighted force-coupled envelope (Env-xF). For springiness, scores for Const-H and Env-xF were significantly higher than those for Const-L. For resistance, the score for Const-H was significantly higher than that for Const-L. These results indicate that springiness during active pressing may be influenced by vibration profiles involving both intensity and temporal variation, whereas resistance may be influenced mainly by the overall vibration intensity. These findings suggest that vibrotactile stimulation may expand the range of haptic expressions without changing the physical reaction force.

Keywords: Medical Informatics, Deep Learning, Image Processing and Pattern Recognition
Abstract: Computed tomography (CT) provides high-resolution volumetric images essential for diagnosing thoracic diseases. However, interpreting 3D CT volumes remains challenging because it requires identifying subtle, localized abnormalities within complex anatomical structures. Medical visual question answering (VQA) has emerged as a promising paradigm for automated reasoning over medical images, and 3D vision-language models (VLMs) have extended this capability to volumetric data. These approaches, however, typically process dense visual tokens extracted from entire CT volumes, which can dilute abnormality-related information among numerous irrelevant tokens. In this work, we address this limitation by proposing an anatomy-guided abnormality token proposal framework for 3D CT VQA. Our approach leverages anatomical masks and anatomy-level abnormality labels to identify candidate abnormality regions within the target anatomy without requiring lesion-level annotations. Features pooled from the selected regions are projected into visual tokens and provided to a large language model (LLM) as explicit visual evidence, encouraging the model to ground its answers in spatially meaningful visual information. Experiments on the RadGenome-Chest CT abnormality VQA task show that our method consistently improves lexical metrics over dense full-volume token baselines while using substantially fewer visual tokens. These results highlight the value of compact, selective token representations for abnormality-centered reasoning in 3D CT VQA.

Keywords: Quantum Cybernetics, Metaheuristic Algorithms, Intelligent Industrial Systems
Abstract: Portfolio optimization is a fundamental problem in financial decision making and has been widely studied using evolutionary and metaheuristic algorithms. However, recent market volatility has increased the demand for low-risk portfolio construction, where downside exposure and asset interactions must be carefully considered. This task remains challenging because the search space grows rapidly with the number of candidate stocks, and low-risk solutions are difficult to identify in dynamic financial landscapes. Although classical metaheuristic methods have shown effectiveness in large-scale optimization, relatively few studies specifically focus on low-risk portfolio construction, and further improvements in convergence efficiency are needed in rapidly changing stock markets. To address these issues, this study proposes a fuzzy-entropy-based quantum-inspired tabu search (FE-QTS) framework for low-risk portfolio optimization. FE-QTS integrates a fuzzy logic controller to adaptively regulate the quantum rotation step size according to search progress, while an entropy-guided diversification mechanism monitors population convergence and injects diversity when stagnation occurs. The search process is further guided by a volatility-oriented objective function with a dynamic loss penalty to discourage loss-making portfolios. Experiments on real-world stock data show that FE-QTS achieves lower risk evaluation values and reduced maximum drawdown compared with conventional metaheuristic and quantum-inspired baselines. These results demonstrate that the proposed framework provides an adaptive and risk-aware approach for robust portfolio construction in volatile markets.

Keywords: Human-Machine Interface
Abstract: n this study, a biological signal-controlled interface that argument body function: a voluntary control method by the electromyogram (EMG) signals of a multi-freedom super numerary robotic finger equivalent to the sixth finger is proposed. Traditionally, it was possible to realize the separation of the EMG pattern corresponding to the sixth finger movement and the original hand motions, but multi-degree-of-freedom control was difficult. In the proposal method, the classification of multiple motions is realized by combining frequency analysis and time series neural networks for multiple sixth finger movements. In the experiment, the EMG signal corresponding to the four motions of the sixth finger was measured, and real-time motion estimation was performed. From the results, it was revealed that although it was difficult to accurately classify four motions, stable classification could be achieved.

Keywords: Trust in Autonomous Systems, Cyber-physical systems, Digital Twin
Abstract: Manufacturing AI systems must autonomously adapt to continuous distributional shift from raw-material variability, ambient changes, and equipment aging, under strict safeguard and operator-trust requirements where model failures risk physical damage. This paper presents a closed-loop Cyber Physical System (CPS) for autonomous model lifecycle management in automotive manufacturing, deployed since 2023. The system manages product-specialized model pairs: a sequence-to-sequence physics model (LPP) serving as a digital twin, and a deep Reinforcement Learning (RL) control policy (LCP) trained against it. Per retraining cycle, multiple model variants spanning architecture families and RL algorithms compete; only the best-scoring candidate advances. A Conductor orchestrator autonomously manages plant-wide model inventories with dependency-aware retraining and Proportional-Integral-Derivative (PID) fallback. Reflecting the principle of Human Centric Intelligence, the LCP composite score embeds an operator trust gate penalizing policies deviating from established practice; without it, 23% of policies are rejected by operators despite passing accuracy thresholds. Across multiple facilities, LCP controlled processes achieve process stability improvements of 28–45% over uncontrolled baselines with zero safety incidents.

Keywords: AI and Applications, Application of Artificial Intelligence, Intelligent Industrial Systems
Abstract: Industrial surface defect segmentation suffers from severe class imbalance: the dominant material occupies most pixels while critical defects appear in only a handful of training instances. Standard plug-and-play attention modules (SE, CBAM, ECA) derive weights from per-sample statistics dominated by frequent patterns, amplifying what is already abundant and leaving rare-class performance largely unchanged. We introduce Defect Archetype Spatial Attention (DASA), a feature-enhancement block combining a fixed morphological prior bank with a learned archetype-conditional channel and spatial gate. DASA is plugged between OneFormer's pixel decoder and its downstream consumers which are the transformer decoder and the mask head so that the high-resolution mask features F₁⁄₄ are enhanced once and feed both query cross-attention and the mask einsum. On a 10-class industrial elastomer defect dataset, DASA improves defect-only mean IoU by +5.70 over the OneFormer + Swin-L baseline (0.503 → 0.560), with double-digit gains on three rare classes (Tool marks +23.0, Mesh +11.4, Spring +10.2). DASA requires no class supervision, no auxiliary loss, and no change to training.

Keywords: Reinforcement Learning, Robotic Systems, Cyber-physical systems
Abstract: Deep reinforcement learning (DRL) is a promising approach for robot path planning in dynamic environments, but conventional training pipelines rely on FP32 arithmetic, automatic differentiation, and floating-point optimizers. This paper studies an integer-only Dueling DQN training method for a dynamic grid-based robot path-planning task on resource-constrained edge AI systems. The proposed method replaces the FP32 update path with fixed-point temporal-difference (TD) target computation, integer TD-error propagation, manual backward updates, per-layer scaling, shift-and-round operations, and saturating discrete weight updates. Preliminary experiments compare random, FP32, and integer-only agents across multiple seeds and analyze replay warm-up, target-network synchronization, and TD-error clipping. The best integer-only setting, with a target synchronization interval of 100 updates and a TD clipping threshold of 1024, increases successful goal-reaching episodes during the last 20 training episodes from 11/100 to 54/100 and reduces the TD clipping rate from 0.4759 to 0.1019. These results indicate that the stability of integer-only DRL depends on both arithmetic precision and RL-specific protocol choices for dynamic path planning.

Keywords: Brain-Computer Interfaces, Human-Machine Interaction, Machine Learning
Abstract: Electrocorticography (ECoG) is a high-fidelity modality for advanced motor brain-computer interface (BCI) applications. However, real-world hand-pose decoding requires systems that continuously monitor the brain activity to distinguish between resting states and specific movements. Previous studies often rely on single-stage flat multiclass classifiers that merge state detection and pose identification, failing to exploit their distinct neurophysiological origins. To address this challenge, we propose a state-aware two-stage guarded-rest architecture that explicitly separates event-related desynchronization (ERD)-driven state detection from high-gamma-driven pose identification. Evaluated using rigorous internal cross-validation (stratified 10-fold for the four-class task and nested for the three-class task) with all preprocessing and model selection confined strictly to training folds, our two-stage model achieved high accuracies of 99.17% in the four-class guarded-rest task and 98.89% in the three-class movement task. Post-hoc kinematic analysis traced the sole error to a mismatch between visual instruction and executed movement, while the ERD-driven state gate correctly identified rest from movement, indicating that state detection remained functionally decoupled from pose identity. The architecture outperforms single-stage baselines, including a flat multiclass classifier (+3.3% gain) and a channel PCA ensemble (+4.2% gain), suggesting that explicit task decomposition yields stronger predictive stability than either single-stage joint classification or blind fusion, while preserving directional consistency, despite low statistical power at high accuracy, pending cross-participant validation.

Keywords: Machine Learning, Big Data Computing, Application of Artificial Intelligence
Abstract: Anomaly detection is a fundamental problem with broad applications. This study focuses on the subspace-based detection strategy, which assumes that normal data lie in a low dimensional subspace whereas anomalies do not. Common approaches to learn the subspace include PCA and matrix factorization, which have limited efficiency as they directly operate on the training matrix. In this paper, we propose an efficient subspace-based anomaly detection approach based on random sketching (RSAD), which learns the subspace from a randomly sampled sketch of the training matrix. Experimental results on multiple benchmark data sets show RSAD consumes substantially less training time than all baselines while maintaining competitive detection performance.

Keywords: Cyber-physical systems, Medical Informatics, Cyber Systems and Engineering
Abstract: Edge machine learning (ML) deployments increasingly rely on per-inference timing measured by software clocks such as Python's perf_counter, but these measurements are not always validated against external hardware references on embedded Linux, and edge ML benchmarking methodologies typically do not isolate platform-dependent instrumentation overhead. This paper reports a preliminary characterization of GPIO call overhead in hardware-validated edge ML inference timing on two embedded platforms running a one-dimensional convolutional neural network (1-D CNN) arrhythmia classifier on electrocardiogram (ECG) data from the MIT-BIH Arrhythmia Database, with five classes per the Association for the Advancement of Medical Instrumentation (AAMI) EC57 standard. Across n = 10 trials on each platform at a controlled steady-state baseline, the per-platform constant on the Jetson Orin Nano (TensorRT FP16, Jetson.GPIO) is approximately −20 µs, and on the Raspberry Pi 4 (ONNX Runtime CPU, pigpio) approximately −86 µs, yielding a cross-platform asymmetry of approximately 66 µs that is large relative to commonly used uniform validation tolerances. The Jetson constant is well-approximated by direct GPIO call duration (the direct profile accounts for ~88% of the platform constant), while the Pi direct profile over-predicts the platform constant by ~19%, motivating empirical per-platform calibration in the deployed measurement context. The Pi constant is not a single sharp value but exhibits a cross-day range of approximately 6 µs across the three sessions sampled, while the Jetson constant reproduces to within approximately 0.14 µs. These preliminary results suggest that cross-platform edge ML timing studies may benefit from platform-aware and potentially session-aware validation gates.

Keywords: Haptic Systems, Assistive Technology, Human-Machine Interaction
Abstract: This paper proposes a vibrotactile feedback method based on acoustic signals generated during object contact to restore tactile sensation in myoelectric prosthetic hands. Contact sounds are recorded using a microphone mounted on the prosthetic fingertip, and prominent frequency components related to surface characteristics are extracted and inverse-transformed to reconstruct vibrations that reflect the contacted surface. The reconstructed signals drive multiple vibration motors placed around the forearm, generating an apparent localized vibration sensation. First, the effects of microphone directivity on the frequency characteristics of contact sounds were evaluated to identify a suitable sensing configuration. Surface roughness discrimination was then compared among the human hand, prosthetic socket, and acoustic-based tactile feedback conditions. The results confirmed that the proposed method enabled participants to distinguish multiple sandpaper grades with performance comparable to that of the human hand.

Keywords: Human-Computer Interaction, Human Factors, Intelligent Internet Systems
Abstract: Voice is a critical component of identity, yet gender-affirming voice training often lacks accessible, personalized guidance. Existing voice training software uses real-time visual feedback to help users perform exercises, but users often struggle to maintain practice over longer time intervals. This paper thus presents a preliminary intelligent coaching module integrated into TruVox, our existing web-based voice training platform. The module includes two functions: exercise recommendation based on voice assessment and intra-exercise target adaptation. Qualitative usability evaluations with six transfeminine users showed a desire for such features and found generally positive impressions, but also provided guidance for future work. In the long term, such coaching features may improve adherence and consequently long-term outcomes of software-aided voice training.

Keywords: Application of Artificial Intelligence, Machine Learning
Abstract: Musculoskeletal injuries remain a concern in industrial and healthcare environments. Devices like robotic exoskeletons can reduce injury risk by aiding the user during lifts, but require knowledge about user movements, characteristics and current state. There are many algorithms to classify such information from sensor data, but no consensus on their relative effectiveness. This work compares different classification algorithms for classification of five lift and user aspects (the weight of the load, how high it is lifted, lifting posture, level of fatigue, and fitness level) using wearable electromyography and inertial sensors. Data were collected from 18 participants performing two repetitions of 24 lift variations. The start and end of each lift were identified in presegmented recordings using a distance-based lift detection algorithm. Features were extracted from sensor signals, reduced via feature selection methods, and used to train different classifiers. Results show the comparative effectiveness of classification methods in different stoop or squat lifting scenarios for different levels of user fitness and fatigue and serve as a step toward better assistive devices and fatigue reduction strategies.

Keywords: Affective Computing, Human-Computer Interaction, Multi-User Interaction
Abstract: In collaborative scenarios involving two participants, task difficulty must be adapted to accommodate differing skill levels. Current adaptation methods for multi-user scenarios primarily rely on task performance, which does not capture users' internal psychological states. This paper investigates automated affect classification in a collaborative scenario using both participants' physiological measurements, including synchrony features that capture covariation between participants' physiological signals. Forty-one pairs played a collaborative game while electrocardiogram, respiration, skin conductance, and two facial electromyograms were recorded. Three classifiers were used to classify interest/enjoyment, perceived competence, and perceived difficulty into three classes. The highest accuracies were 60% for interest/enjoyment, 68.1% for perceived competence, and 46.1% for perceived difficulty (all achieved using random forests), which is comparable to results from competitive multi-user studies. However, including physiological synchrony features had only minor effects on classification accuracy, so it is unclear whether it justifies the additional calculation. Nonetheless, these findings support the feasibility of physiology-based affect classification in collaborative settings as a step toward dynamic difficulty adaptation.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Computational Intelligence
Abstract: Maintaining a person's identity through occlusion remains a challenging problem in video understanding, especially in human-centered environments where reliable continuity is more important than isolated per-frame detections. Existing pipelines often perform well when people are clearly visible, but struggle when a person is partially hidden, fully occluded, or re-enters the scene after a temporary disappearance. This work presents a human-centric pipeline for occlusion-aware person identity persistence that combines open-set person localization, promptable person segmentation, point-based temporal continuity, and state-memory reasoning. The proposed system uses Grounding DINO for person-level localization, the Segment Anything Model for precise person masks, CoTracker3 for tracking selected visual points through occlusion, and a buffered state engine to manage transitions among visible, occluded, reappeared, and lost states. Unlike box-only tracking pipelines, the proposed architecture treats the segmented person region as the primary visual representation for tracking the same person and uses state-memory logic to support identity recovery after disappearance. The pipeline is evaluated on the Multiple-Object Tracking 2017 benchmark, and results are reported using Higher Order Tracking Accuracy, Association Accuracy, Identity F1 Score, identity switches, and related metrics. Early results show that the architecture is promising, particularly in integrating segmentation and occlusion-aware point continuity; however, reappearance association and identity-switch reduction remain the main open challenges. This work is positioned as an early-stage human-centric perception pipeline for person continuity in dynamic visual environments.

Keywords: Human Factors, Human-Centered Transportation, Wearable Computing
Abstract: Reverse parking is a common yet complex maneuver that poses significant safety risks, particularly for novice drivers, and is therefore important for driver education. This study investigates how driving experience affects gaze behavior and head movement intensity during left-to-right reverse parking using wearable eye-tracking glasses. Data were collected from 100 driving sessions, including 60 sessions from experienced drivers and 40 sessions from novice drivers. Four gaze-related metrics were extracted from four predefined Areas of Interest (AOIs), and head movement intensity was quantified using accelerometer data. Stepwise regression analysis identified visit count, fixation count, and total fixation duration on the left door mirror as significant predictors of head movement intensity. To examine group differences, Mann–Whitney U tests were conducted. Results showed that novice drivers exhibited a significantly higher visit count on the left door mirror than experienced drivers, whereas fixation count and total fixation duration were not significantly different. Notably, experienced drivers demonstrated greater head movement intensity, suggesting a broader and more proactive scanning strategy during the maneuver. In contrast, novice drivers tended to visit the left door mirror more frequently with more limited head movement. These findings highlight the dynamic relationship between gaze behavior and driving experience, offering practical insights for driver education and adaptive monitoring systems aimed at enhancing safety during low-speed maneuvers.

Keywords:
Abstract: Quantum Machine Learning (QML) and Quantum Artificial Intelligence (QAI) are emerging domains poised to bring breakthroughs in computational science. However, significant barriers to adoption remain, including the complexity of model design, the arduous process of hyperparameter tuning, and the need for robust governance. This article introduces Quoptuna, an open-source Python toolkit designed to simplify and accelerate the entire QML workflow. Quoptuna provides an integrated platform for automated model selection, hyperparameter optimization, experiment tracking, and explainable AI (XAI). By leveraging classical simulators, it lowers the barrier to entry, allowing for efficient benchmarking without requiring immediate access to physical quantum hardware. Furthermore, Quoptuna introduces a novel capability for generating reports using large language models (LLMs), providing automated, governance-oriented insights into model performance, fairness, and transparency. This work aims to democratize QML, making it more accessible to a broad audience of researchers, engineers, and developers, thereby accelerating the journey from research to real-world applications.

Keywords:
Abstract: Graph Neural Networks (GNNs) are effective for graph-structured data, yet often degrade under scarce labels, noisy graphs, or complex relational patterns. Existing graph-based boosting methods typically rely on external boosters or explicit sample reweighting, which may disrupt end-to-end optimization and weaken GNNs' relational inductive biases. In this paper, accepted to IEEE Transactions on Computational Social Systems (TCSS) 2026, we propose Adaptive-boosting Graph Neural Networks (AdaGNN), a graph-native boosting framework for robust graph learning. AdaGNN sequentially trains GNN learners using progressively refined node embeddings. After each round, correctly classified nodes are frozen, enabling later learners to focus on remaining errors. This Embedding Freeze mechanism implicitly propagates error signals through graph topology without explicit reweighting or multi-stage coordination. Experiments show that AdaGNN consistently improves node classification accuracy and robustness, especially in low-label scenarios.

Keywords: Homeland Security, Information Assurance and Intelligence, Systems Safety and Security
Abstract: Cyber Threat Intelligence (CTI) sharing is a fundamental component of collaborative cybersecurity in governmental environments. However, improving the integrity, traceability, and trustworthiness of shared incident data across decentralized entities remains a challenge. Existing approaches often lack a unified mechanism that integrates automated data immutability with expert-driven quality assessment. To address this gap, this paper proposes BlockCTI, a blockchain-based architecture that integrates a Proof of Quality (PoQ) consensus mechanism with the NATO Admiralty Scale. Developed through a systems engineering methodology, the proposed architecture specifies a decentralized validation process that incorporates human expertise into automated security workflows using Structured Threat Information eXpression (STIX), Trusted Automated eXchange of Intelligence Information (TAXII), and Malware Information Sharing Platform (MISP) standards. This study provides conceptual modeling and architectural design for a transparent, trust-aware ecosystem designed to enhance situational awareness and to support informed, risk-based decision-making by integrating data integrity and expert validation natively.

Keywords:
Abstract: A robust understanding of a human’s internal state can greatly improve human-robot teaming, as estimating the human teammates’ workload can inform more dynamic robot adaptations. Existing workload estimation methods use standard machine learning techniques to model the relationships between physiological metrics and workload. However, such methods are not sufficient for adaptive systems, as standard machine learning techniques struggle to make accurate workload estimates when the human-robot team performs unknown tasks. A meta-learning based workload estimation algorithm is introduced and an initial analysis is conducted to show how adapting a machine learning model’s parameters using task-specific information can improve result in more accurate workload estimates for unknown tasks.

Keywords: Adaptive Systems, Sustainable Systems
Abstract: ISO/IEC/IEEE 16085 promotes risk management throughout Systems Engineering (SE) lifecycle process and is an important milestone to assure successful SE outcomes. On the other hand, for some methods included from traditional and generic risk management domains, they have variants that are better customized for SE tasks. This paper reports the recent progress of proposals of updates to this standard to include risk management methods that are closely customized for SE tasks.

Keywords: Digital Twin, Smart Buildings, Smart Cities and Infrastructures, Social simulations
Abstract: In modern manufacturing, a simulation based analytical approach is essential for implementing Cyber-Physical Production Systems (CPPS). Recently, research utilizing NVIDIA's Isaac Sim, which provides high-performance physics engine based Digital Twin (DT) and simulation, has been actively conducted. However, the process of configuring complex physical parameters during the initial environment deployment phase remains limited by its heavy reliance on engineers' manual trial-and-error, leading to significant time and resource consumption. To address this limitation, this study proposes the RFR-DE framework, which integrates simulation automation for efficient physical parameter exploration, machine learning based physical behavior prediction, and an optimization algorithm driven parameter estimation procedure. The proposed method was validated on a single-case robotic control system, demonstrating a practical advantage over conventional manual trial-and-error approaches.

Keywords: Social simulations, Agent-Based Modeling, Heuristic Algorithms
Abstract: Recently real-scale simulations (RSSs) have attentions from researchers. Since RSSs are targeting specific regions, detailed population data at real scale is essential. While local governments maintain resident registers based on residency records, the use of those records is highly restrictive due to privacy reason and is difficult to use them for simulations. To establish a research infrastructure for simulations using real-scale population data, several research projects prepare synthetic populations for some countries. Murata et al. have added workplace attributes to syn- thetic population data. In their study, workers were assigned independently at the municipal level based on their place of residence. However, no consideration was given to whether the results of assigning workers from multiple municipalities to workplaces accurately captured the statistical character- istics of the municipalities where they actually worked. This study mitigates the inconsistencies on the place of employment caused by the independent assignment based on the place of residence by Murata et al., we propose a method that simultaneously determines and assigns the municipalities where workers are employed nationwide using Iterative Proportional Fitting (IPF). The proposed method improves the number of workers who are mistakenly assigned.

Keywords: Biometric Systems and Bioinformatics, Application of Artificial Intelligence, Reinforcement Learning
Abstract: Long-horizon cotton irrigation must remain safe under forecast errors, degraded observations, and abrupt water-budget changes. Many continuous reinforcement-learning controllers map the state directly to a daily irrigation amount without explicitly determining when a management mode should persist, switch, or be verified. We formulate an execution feedback memory-driven semi-Markov architecture that sepa rates a high-level language agent from a shared, risk-conditioned continuous deep reinforcement-learning (DRL) worker. At event triggered option boundaries, the agent selects continue, switch, or verify; the worker alone produces the daily irrigation amount. A deterministic residual module records structured failure attributions in memory, while a digital twin serves only as a cost-aware verification tool. We define a falsifiable out-of distribution evaluation protocol for weather, sensing, and budget disturbances to assess whether persistent option management can improve irrigation safety and robustness while controlling verification cost.

Keywords: Mechatronics, AI and Applications, Connected and Automated Vehicles
Abstract: The global agricultural sector is undergoing a profound transformation driven by the dual imperatives of decarbonization and digitalization. Conventional diesel-powered agricultural machinery, while historically indispensable for modern food production, contributes significantly to greenhouse gas emissions, noise pollution, and escalating operational costs. Against this backdrop, the transition toward new-energy electric agricultural machinery (EAM) has emerged as a strategically critical pathway for achieving sustainable and intelligent farming systems. Central to the viability and performance of any EAM is the Vehicle Control Unit (VCU)—the electronic "brain" that orchestrates all powertrain components, sensors, and actuators in a coordinated and efficient manner. This paper presents a comprehensive account of the design, implementation, and experimental validation of a dedicated VCU developed for a multi-motor electric crawler-type rotary tiller—a representative class of agricultural equipment widely employed in both dryland and paddy field operations across Asia and beyond. The proposed VCU architecture is built upon a hierarchical control framework that establishes clear functional boundaries and communication protocols among the three principal subsystems of the machine: the traction drive system, which comprises two independently controlled permanent magnet synchronous motors (PMSMs) driving the left and right crawler tracks; the power take-off (PTO) implement drive system, which is powered by a dedicated PMSM coupled directly to the rotary tillage blade shaft; and the autonomous navigation system, which integrates a Real-Time Kinematic Global Navigation Satellite System (RTKGNSS) receiver, an inertial measurement unit (IMU), and wheel encoders for precise state estimation. At the hardware level, the VCU is implemented on an automotive-grade multi-core microcontroller with multiple CAN FD interfaces, analog and digital I/O channels, and Ethernet connectivity for high-bandwidth data logging. The software architecture follows a modular, layered design that separates low-level device drivers, middleware for inter-process communication, and high-level control algorithms, thereby facilitating maintainabi

Keywords: Reinforcement Learning, Evolutionary Computation, Game Theory and Multi-Agent Systems
Abstract: Multi-agent reinforcement learning (MARL) algorithms have shown great potential in various domains. However, MARL is vulnerable to adversarial attacks where compromised agents execute arbitrary worst-case actions to disrupt coordination. Existing robust MARL methods often focus on defending a single attack pattern, thereby struggling against diverse attack modes. To address this problem, we propose the Multi-Objective Memetic Attack (MOMA) framework, an adversarial attack method inspired by Evolutionary Algorithms (EAs) to generate highly diverse and destructive adversarial attacks for robust MARL. Unlike standard adversarial attack methods that only learn a single attack pattern, MOMA formulates the attack process as a multi-objective optimization problem, and uses the multi-objective evolutionary algorithm (MOEA) to optimize a population of different attack strategies. MOMA innovatively utilizes the behavioral distance of each individual to the population consensus as the metric to assess the behavioral novelty of attackers, effectively preserving distinct attack modes within the attacker population. Through the integration of the MOEA and an advanced memetic update mechanism, MOMA can efficiently explore potential vulnerabilities within the MARL policies. Based on MOMA, we further propose MOMA-AT (MOMA-based Adversarial Training), an adversarial training paradigm utilizing the discovered diverse attack strategies learned by MOMA to enhance the robustness of the multi-agent system against varying threat models. Empirical results demonstrate the impact of MOMA and the significant robustness improvement under MOMA-AT.

Keywords: Reinforcement Learning, Game Theory and Multi-Agent Systems, Neural Networks and their Applications
Abstract: Multi-Agent Reinforcement Learning (MARL) has shown broad success across various domains. However, existing MARL algorithms often lack robustness on observations and are sensitive to small perturbations, limiting their practical deployment in real-world scenarios. Previous studies have shown the vulnerability of MARL algorithms through adversarial attacks, but existing robustness-enhancing methods often lead to trade-offs during training, negatively affecting performance in normal (non-adversarial) environments. In this work, we aim to improve the robustness of MARL algorithms in adversarial settings without sacrificing performance in normal environments. We first formulate the adversarial attack on MARL as an Adversarial Observation Dec-POMDP (AO-Dec-POMDP) to investigate the properties of the problem, then propose a Pareto-Robust Optimization scheme for MARL (PRO-MARL) with theoretical guarantees, which can be integrated with various MARL algorithms. Experiments on SMAC demonstrate the superior robustness of PRO-MARL in adversarial environments while maintaining outstanding performance in normal settings.

Keywords: Brain-Computer Interfaces, Affective Computing, Application of Artificial Intelligence
Abstract: Graph Neural Networks (GNNs) have been widely used to capture spatial functional connectivity patterns to improve electroencephalography (EEG)-based depression recognition performance. However, the functional connectivity of brain networks in patients with depression exhibits an inherent hierarchical structure, making it difficult to capture accurate connection patterns. To address these issues, this paper proposes a novel model named Sample-Adaptive Hyperbolic Graph Neural Network (SA-HGNN), which aims to accurately extract the authentic hierarchical structure of depression-affected brain networks. Specifically, the proposed model comprises three core modules. First, a Sample-Adaptive Graph Construction module dynamically constructs personalized brain network topologies to capture more complex spatial relationships within the brain network. Second, hyperbolic graph convolution is employed to overcome the representation bottlenecks of Euclidean space, leveraging hyperbolic geometry to precisely capture latent hierarchical relationships within the brain network. Finally, an Attention Pooling module adaptively filters out highly redundant noise channels in EEG signals, effectively mitigating the interference of inherent noise on the authentic hierarchical topology. Extensive experiments on public EEG datasets demonstrate the superior performance of our method across resting-state and task-related paradigms, validating its robustness to noise and efficacy in capturing abnormal functional connectivity patterns in brain networks of patients with depression.

Index Terms--Major Depressive Disorder Recognition, Electroencephalography, Graph Neural Networks, Hyperbolic Geometry, Graph Pooling.

Keywords: Brain-Computer Interfaces, BMI Emerging Applications, Human-Machine Interaction
Abstract: Brain–computer interface (BCI)-assisted exoskeletons have the potential to improve lower-limb stroke rehabilitation by translating user intention into rehabilitation- oriented robotic assistance. However, lower-limb rehabilitation requires not only gait initiation but also reliable control over functional transitions such as standing up, walking, and sitting down. This paper presents a three-stage wearable augmented reality and brain–computer interface (WAR-BCI) control framework for a lower-limb exoskeleton for stroke rehabilitation. The proposed framework integrates wireless electroencephalography (EEG), AR-based visual guidance, and stage-specific EEG decoding to generate control commands for three sequential rehabilitation actions: sit-to-stand, walking, and stand-to-sit. Specifically, frontal theta-band modulation is used for standing control, movement-related cortical potentials (MRCPs) are used for walking control, and occipital alpha-band modulation is used for sitting control. Online experiments involving healthy participants and stroke patients showed high training performance and encouraging online control performance for the complete three-stage task sequence. The results support the feasibility of the proposed WAR-BCI control strategy for lower-limb exoskeleton-assisted stroke rehabilitation.

Keywords: Medical Informatics, Deep Learning, Brain-Computer Interfaces
Abstract: 近红外光谱（fNIRS）受限于 联合建模局部血流动动力学的难度 异常现象和全球时空依赖性。前往 针对这个问题，我们提出fNIRS-CTPDNet， 神经生理学引导的CNN-变换器融合框架 利用fNIRS信号进行PD分类。所提模型 遵循本地到全球设计。时间上、深度上，以及 空间卷积用于编码延迟血流动动力学 动力学、通道特异性皮层反应， 分别是短距离通道间协调。A 具有保持信道分词的变换器编码器为 进一步引入以捕捉长距离跨区域和 在保持信道语义的同时实现跨时间依赖关系。 通过将神经生理先验纳入局部 特征提取和全局依赖学习，以及 拟议框架超越了简单的泛泛 CNN-变压器堆栈，学习更具判别性 PD识别的表示。实验 涉及手指敲击、行走的公共PD-fNIRS数据集， 静止状态

Keywords: Brain-Computer Interfaces, Deep Learning, Transfer Learning
Abstract: Abstract— Functional near-infrared spectroscopy (fNIRS) is a promising modality for autism spectrum disorder (ASD) clas- sification, yet existing approaches assume temporally aligned evaluation. In practice, the optimal observation window varies across subjects due to differences in hemodynamic delay and neurovascular coupling, creating a temporal distribution shift that degrades performance. We formalize this as a cross-time- window transfer problem, introducing a protocol that varies window length (2.5–10 s) and offset within biological motion trials. Using topographic map representations of fNIRS record- ings, we benchmark three vision architectures under two zero- shot baselines and eight adaptation strategies under leave-one- subject-out cross-validation (N=124). Key findings: (1) zero- shot cross-window accuracy is near chance (54–69%); (2) ≈5% subject-specific fine-tuning recovers 90–96%, while a subject- specific upper bound reaches 97–100%, identifying inter-subject variability as the dominant barrier; (3) domain-adversarial and self-supervised strategies achieve 78–90% without target- subject data; and (4) discriminative information is recoverable from windows as short as 2.5 s. These findings provide a practical roadmap for deploying fNIRS-based ASD classifiers under realistic temporal variability.

Keywords: BMI Emerging Applications, Hybrid Models of Neural Networks, Fuzzy Systems, and Evolutionary Computing, Machine Learning
Abstract: This study proposes a hybrid machine learning framework to estimate quantified impulsivity scores in children using task-related EEG signals. By focusing on neural dynamics during the Continuous Performance Test (CPT), we aim to evaluate the feasibility of objective impulsivity assessment across clinical and typically developing populations. EEG data were acquired from 107 children, including 72 participants with ADHD and 35 typically developing (TD) peers, during the CPT. To address the inherent variability in pediatric physiological data, a feature selection pipeline using minimum Redundancy Maximum Relevance (mRMR) was implemented. We evaluated a Hybrid Top-6 modeling strategy (integrating Bagging and Multilayer Perceptron architectures) against standalone SVR and MLP models to predict impulsivity scores based on CPT performance indices. he hybrid framework demonstrated higher predictive consistency than standalone regressors. In the TD group (n=35), the system achieved its highest performance with 75–80% accuracy in score estimation. For the ADHD group (n=72), although the clinical heterogeneity presented greater challenges, the hybrid approach stabilized predictive outcomes and maintained a robust baseline compared to individual models. Key predictors identified via mRMR included prefrontal TBR (Fp1, Fp2) and occipital alpha (O1, O2), confirming the role of frontal-occipital networks in inhibitory control during task engagement. Our findings suggest that task-based EEG, when processed through a hybrid machine learning architecture, provides a reliable tool for quantifying impulsivity. This framework effectively manages the variability in pediatric clinical samples, offering an objective foundation for individualized treatment monitoring in ADHD.

Keywords: Autonomous Vehicle, Environmental Sensing, Deep Learning
Abstract: Multi-view perception based on Bird's-Eye-View (BEV) representation plays a crucial role in scene understanding for autonomous driving. However, missing camera views caused by occlusions or sensor failures often lead to incomplete perception, degraded spatial consistency, and reduced detection accuracy. To address these challenges, we propose DiT-MVR, a scalable Diffusion Transformer (DiT)-based framework for multi-view feature reconstruction. Unlike discriminative regression methods such as M-BEV, DiT-MVR leverages generative diffusion modeling to restore missing-view features. By integrating Adaptive Layer Normalization modulation and view-specific context processing, our method achieves stronger global dependency modeling and higher fidelity feature recovery. Experiments on the nuScenes dataset demonstrate that DiT-MVR delivers significant improvements under various missing-view conditions. For example, in the front-left missing scenario, DiT-MVR surpasses M-BEV by 1.7% NDS, highlighting its robustness and strong potential for reliable BEV perception in real-world autonomous driving systems.

Keywords: System Modeling and Control, Robotic Systems
Abstract: This paper proposes an adaptive-tick behavior tree (@BT) that balances reactivity and efficiency in dynamic environments. Currently, BTs are mainly time-driven or event-driven. Time-driven BTs waste resources on unnecessary ticks, while event-driven BTs rely on a task-specific event list. @BT addresses both issues by integrating H∞ controllers into action nodes to compute control inputs and monitor the input error. When the error exceeds a dynamic threshold, the controller publishes a tick event to the event queue. The BT engine executes a tick after detecting the event. We further improve computation efficiency by integrating an early-exit mechanism in the child action nodes of control flow nodes with memory, so that expensive high-level algorithms are invoked only when necessary. Simulation experiments show that @BT significantly reduces the tick frequency while maintaining task performance.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Fine-grained object detection in complex scenes requires subtle local textures, long-range sparse cues and low-cost inference. Large-kernel convolutional backbones and transformer-based real-time detectors each address part of this problem, but are rarely co-optimised. We propose DC-Net, a unified framework that couples a dilated-reparameterised backbone, a dual semantic-spatial calibration neck and a DETR-style query decoder. The DR-ELAN backbone trains parallel dilated small kernels and reparameterises them into a single equivalent large-kernel convolution at inference. The DCSM neck cascades Context Feature Calibration (CFC) and Spatial Feature Calibration (SFC) to align multi-scale features before the query decoder. We evaluate DC-Net on PASCAL VOC, MS COCO, and on Snake-27, a fine-grained stress-test dataset of 27 snake species we curate for camouflage, occlusion and high intra-class similarity. DC-Net reaches 81.2% mAP50 on VOC, 61.6% mAP50 on COCO, and 95.1% mAP50 on Snake-27 while running at 46.5 FPS with only 12.7M parameters and 31.5 GFLOPs, matching or surpassing recent real-time transformer and YOLO-series detectors at a comparable parameter budget. Controlled ablations confirm that dilated reparameterisation consistently outperforms vanilla large-kernel and single-dilation baselines at the same inference cost.

Keywords: Machine Vision, Machine Learning, Image Processing and Pattern Recognition
Abstract: Vision-language (VL) tracking, which incorporates natural language for robust target localization, remains underexplored in aerial trackers. Existing VL trackers often neglect the misalignment between static language descriptions and targets’ dynamic states, a problem particularly acute in complex UAV environments. To this end, we propose ATTSTrack, a novel VL aerial tracker designed to dynamically align linguistic semantics with visual representations through an adaptive textual tokens selection mechanism. Specifically, we employ a cross-modal attention module to estimate tokens relevance and a gating mechanism to select the most discriminative text tokens. This enables the language input to accurately reflect the target’s current state and adapt to dynamic changes in the scene. Furthermore, we introduce language-adaptive feature modulation, which leverages the refined textual information to guide the generation of more discriminative visual features. Extensive experiments on two public aerial vision-language tracking benchmarks demonstrate that our method achieves state-of-the-art performance, validating its effectiveness and robustness.

Keywords: Cognitive Computing, Representation Learning
Abstract: Cognitive intelligence relies on efficient memory for deep knowledge understanding and complex logical inference. However, existing models suffer from three critical bottlenecks: 1) Euclidean space-based memory modeling brings dimensional inefficiency and flawed hierarchical association modeling; 2) lack of biologically plausible dynamic memory regulation fails to mimic human memory's decay, consolidation and active forgetting; 3) unfocused inference causes redundant computation and reduced reasoning precision. To solve these issues, we propose HMAI, a cognitive-inspired hyperbolic memory-augmented inference framework with three core innovations: a dual-layer dynamic memory update mechanism simulating human memory dynamics; a hyperbolic graph memory augmentation paradigm enabling low-dimensional hierarchical representation and breaking Euclidean topology limitations; a routing-constrained focused inference method screening key memory units for efficient and accurate reasoning. Experiments on representative visual and textual benchmarks show that HMAI achieves superior performance compared to baseline models.

Keywords: Connected and Automated Vehicles, Systems Safety and Security, Deep Learning
Abstract: The reliability of in-vehicle intrusion detection systems is critical to the safety of intelligent connected vehicles. Existing supervised methods mostly rely on closed-set assumptions and cannot recognize zero-day unknown attacks. Anomaly-based approaches enable unknown attack detection but typically suffer from high false negatives and limited attack-type discrimination. Modeling in-vehicle intrusion detection as open-set recognition is thus more practical, yet two key challenges remain. First, severe class imbalance leads unknown attacks to be misclassified as majority classes and tail classes to be incorrectly rejected as unknown. Meanwhile, high inter-dataset attack overlap and inconsistent cross-vehicle data base container (DBC) files hinder the transfer of outlier exposure (OE) methods, which rely on external auxiliary data to disentangle out-of-distribution class spaces. Second, conventional feature extractors fail to adequately model complex temporal patterns and diverse attack morphologies, resulting in unstable performance over varying unknown classes. To address these issues, we propose PMC-SPOR, a Proto-Mahalanobis Contrastive Learning framework with soft-guided pseudo-OOD repulsion. It leverages dual-stream mixed-sample routing to construct soft-label supervision and pseudo-OOD constraints without external data, alleviating long-tail bias while enhancing unknown rejection. We further design a multi-scale chunked spatio-temporal attention module to capture both local CAN patterns and long-range dependencies, improving stability in open-set recognition. Experiments show that PMC-SPOR outperforms state-of-the-art baselines, especially on long-tailed and complex in-vehicle datasets.

Keywords: Neural Networks and their Applications, Image Processing and Pattern Recognition, Computational Life Science
Abstract: Accurate multimodal brain tumor classification is critical for early diagnosis and clinical decision support. CT and MRI provide complementary information, yet existing lightweight methods still show three practical limitations: subtle lesion cues are easily diluted by dominant background responses, cross-layer feature reuse is often fixed rather than input-adaptive, and spatial positional priors are insufficiently exploited in multi-scale fusion. To address these issues, we propose TransXNet-DMC, which follows a coordinated enhancement pipeline: Differential Attention (DA) first suppresses common-mode attention noise to highlight subtle discriminative cues, Multi-path Dynamic Dense Connection (MUDD) then performs adaptive historical-feature routing for robust cross-layer interaction, and Multi-scale Coordinate Attention (MCA) finally refines lesion-aware representation with joint scale and coordinate modeling. On a paired CT-MRI brain tumor dataset, TransXNet-DMC achieves 97.45% accuracy and 0.9951 AUC, with consistently strong precision, recall, and F1-score. The code is available at https://github.com/MissuForever/TransXNet-DMC.

Keywords: Intelligent Transportation Systems, Connected and Automated Vehicles, Reinforcement Learning
Abstract: Existing pseudonym change schemes in low-density traffic environments fail to protect vehicle location privacy due to the lack of an effective and anti-detectable chaff message generation mechanism. To address this issue, we propose AntiDet-Traj, an anti-detectable trajectory generation method for location privacy that overcomes the limitations of existing trajectory generation methods. By leveraging the Proximal Policy Optimization (PPO) algorithm combined with Generalized Advantage Estimation (GAE), along with a fine-grained reward mechanism designed in this work, AntiDet-Traj can efficiently generate chaff trajectories between two specific points that are hard for attackers to detect. When AntiDet-Traj is applied to pseudonym change schemes, it effectively mitigates location privacy leakage in low-density traffic environments by generating anti-detectable chaff trajectories between mix zones. Experimental results demonstrate that AntiDet-Traj achieves higher generation efficiency and higher trajectory quality compared to the state-of-the-art (SOTA) method, reaching a 100.00% success rate, requiring 84.21% fewer training episodes to exceed a 90.00% success rate, and reducing the average time steps and generation time by 41.99% and 72.59%, respectively. Furthermore, the chaff trajectories generated by AntiDet-Traj exhibit strong anti-detectability under both full-feature and single-feature settings, reducing the Recall of full-feature-based detection by 58.06%-80.37% compared to the SOTA method and 88.60%-93.51% compared to the baseline method.

Keywords: Medical Informatics, Image Processing and Pattern Recognition, Deep Learning
Abstract: Coronary artery disease remains a major cause of mortality and disability worldwide. X-ray coronary angiography is the clinical gold standard for localizing stenotic lesions and assessing their severity. However, automated coronary artery stenosis detection still faces three major challenges: the scarcity of high-quality annotated datasets, large vascular scale variation with strong noise interference in angiographic images, and the high computational cost of existing detection models.

To address these issues, we propose a lightweight coronary artery stenosis detection framework, termed Lightweight Vascular Structure Detection Network (LVSD-Net). The proposed framework includes a backbone, VesselSpectFocusNet BackBone, which performs multi-scale vascular feature extraction and incorporates a frequency-adaptive analysis module to enhance fine-grained pathological representation, thereby improving the detection of small stenotic regions. In addition, a lightweight structure-enhanced detection head with shared features is designed to preserve discriminative vascular details while reducing computational complexity. We also construct a high-quality coronary artery stenosis dataset, CASDI, with both lesion localization annotations and stenosis severity labels.

Experimental results show that, compared with YOLOv12, LVSD-Net improves the F1-score and mAP50 by 8.1% and 4.7% on the CASDI dataset, respectively, and by 2.3% and 1.1% on the CADICA dataset. These results demonstrate that LVSD-Net achieves a favorable balance between detection accuracy and computational efficiency, showing strong potential for practical clinical deployment.

Keywords: Machine Vision, Image Processing and Pattern Recognition, Deep Learning
Abstract: Scenario-specific datasets are essential for advancing computer vision in real-world environments. However, public datasets and benchmarks for livestock farming remain limited, which restricts progress in robust visual perception for precision livestock farming. In practice, livestock detection is still challenging due to heavy occlusion, dense animal distribution, and large illumination variations across on-site farming scenes. To address this problem, we introduce RawLivestock, a large-scale real-world dataset for livestock detection. RawLivestock contains 71,000 images of cattle, sheep, and pigs collected from diverse farming environments. Among them, 10,000 images are manually annotated with 17,792 bounding boxes, while the remaining 61,000 images are kept unlabeled to support semi-supervised learning. Based on this data split, the annotated subset is used for supervised training and evaluation, and the unlabeled subset provides additional training data for scalable semi-supervised learning. We further establish a systematic benchmark by evaluating multiple state-of-the-art object detection methods on RawLivestock. The proposed dataset and benchmark provide a challenging testbed for livestock detection and can support future research on intelligent multimedia analysis in livestock farming. The dataset is publicly available at url{https://github.com/ZhaoJuan0304/RawLivestock-Dataset}.

Keywords: Cyber-physical systems, Application of Artificial Intelligence, System Modeling and Control
Abstract: Falsification of hybrid dynamical systems remains challenging due to mode-dependent dynamics and discrete transitions. In this work, we propose a surrogate-based falsification approach enabled by learning a differentiable hybrid automaton model from data. This extends previous surrogate-based falsification methods, which were limited to modeling purely continuous dynamics. Specifically, we employ neural hybrid automata to learn both a latent mode encoder and the corresponding mode-conditioned vector fields. Subsequently, transition guards are inferred from the same trajectory data. We embed the resulting surrogate in a gradient-based optimal control formulation that minimizes a smooth robustness approximation of the safety specification to search for violations, and validate the solution on the original system to ensure soundness. The proposed method consistently uncovers counterexamples in 4 out of 6 of the ARCH-COMP 2024 benchmark specifications; in these cases, it achieves sample efficiency that is competitive with, or higher than, other state-of-the-art tools from the benchmark.

Keywords: Deep Learning, Machine Vision, AI and Applications
Abstract: Most open-vocabulary semantic segmentation (OVSS) methods leverage CLIP’s vision-language alignment to classify each pixel against arbitrary text-defined categories without task-specific training. However, CLIP is trained at the image level, and its local patch features are inevitably contaminated by global semantics. This contamination cascades through self-attention across layers, producing outlier tokens, progressively degrading local discriminability and leaving a pervasive global bias that blurs inter-class boundaries. These effects are amplified in remote sensing imagery by large-scale variations and complex backgrounds. To suppress feature contamination, we propose CLIP-Sanitizer, a training-free pipeline with three concise operations: Feature Rectification (FR) corrects intermediate outlier tokens with local neighbors, Global Debiasing (GD) adaptively subtracts residual [CLS] bias from patch embeddings, and Guided Feature Refinement (GFR) upsamples and filters features to sharpen boundaries. The pipeline introduces no learnable parameters and works with any frozen CLIP Vision Transformer (ViT) backbone. Experiments on ten remote sensing datasets covering satellite, aerial, and unmanned aerial vehicle (UAV) imagery show that CLIP-Sanitizer consistently improves mean Intersection-over-Union (mIoU) over existing training-free methods across diverse remote sensing scenarios.

Keywords: Cooperative Work in Design, Decision Support Systems, Visual Analytics, Communication
Abstract: Design review is a collaborative human-machine systems problem: reviewers comment on different regions of an interface, raise partially overlapping concerns, and must eventually converge on design decisions. Existing datasets and tools support UI critique generation, UI mining, argument mining, and discussion visualization, but rarely connect region-grounded UI comments with disagreement detection and consensus-oriented interaction. This paper presents ConflictVis, an interactive system for collaborative UI design that detects potential disagreement among design comments, groups comments around shared review issues, and visualizes the current status of discussion for collaborative decision support. We construct a region-grounded critique dataset by expanding UICrit comments and aligning them with RICO screenshots, RICO hierarchy files, and RICO Semantics annotations. ConflictVis first groups comments by issue and selects candidate pairs, then estimates comment relations and conflict types before summarizing issue status in a three-panel interface. On 11,430 UI critique comments covering 1,000 mobile screens, the system produces 13,696 candidate comment pairs and 5,158 issue-level review summaries. Evaluation on the constructed benchmark shows macro-F1 scores of 0.724 for issue grouping, 0.571 for comment-relation prediction, 0.428 for conflict categorization, and 0.890 for issue-status estimation. These results indicate that region-grounded issue analysis can make disagreements easier to inspect during collaborative design review. The implementation is publicly available at https://github.com/ggkong/ConflictVis.

Keywords: Systems Safety and Security, Machine Learning, Distributed Intelligent Systems
Abstract: Federated learning (FL) is vulnerable to backdoor attacks, especially under strong non-IID heterogeneity, where benign client updates are highly dispersed and hard-filtering defenses may discard informative benign updates. We propose T-CRSA, a utility-aware backdoor defense that replaces hard exclusion with continuous, temporally smoothed aggregation weights. T-CRSA uses an enhanced update representation, dominant-anchor-based risk estimation, temperature-controlled soft weighting, and exponential moving average (EMA) smoothing. This design targets heterogeneous settings where benign and malicious updates partially overlap and hard decisions become unreliable. Experiments under multiple attack and heterogeneity settings show that T-CRSA preserves clean-task utility better than hard-clustering baselines while maintaining a competitive utility-security trade-off. Under extreme heterogeneity (α = 0.1), T-CRSA achieves 81.33% main-task accuracy and 27.55% backdoor accuracy, avoiding the severe utility collapse observed in the hard-clustering baseline.

Keywords: Generative AI, Image Processing and Pattern Recognition, Neural Networks and their Applications
Abstract: The objective of text-to-image (T2I) personalization is to generate images that faithfully reflect both a user-provided concept and a given text description. While existing plug-and-play frameworks such as DreamMatcher effectively preserves appearance characteristics of a specific concept while maintaining consistency with the text prompt, they struggle to achieve stylized T2I personalization, as the direct injection of value (V) features from the reconstruction process tends to overwrite the stylistic information encoded in the generation path. To address this limitation, we propose a triple-path diffusion architecture that explicitly disentangles and fuses appearance and style representations. Specifically, a novel fusion path is introduced to integrate the spatial structure of the reference image with the style semantics of the text prompt, producing a fused value V that preserves stylistic integrity in the generation process. Moreover, a dual semantic matching guidance mechanism is designed to refine predicted noise across denoising steps, thereby stabilizing feature fusion and ensuring faithful stylized synthesis. Extensive experiments demonstrate that our method not only enables high-quality stylized T2I personalization but also considerably improves prompt fidelity while maintaining subject structural fidelity, remarkably outperforming competitive baselines across both standard and challenging datasets.

Keywords: Medical Informatics
Abstract: Coronary collateral circulation (CCC) is a crucial protective mechanism for patients with chronic total occlusion (CTO). Clinically, the Rentrop grading system is the gold standard for assessing CCC function. However, automated intelligent assessment in this field is currently lacking, with a notable absence of dedicated research and public datasets. Existing clinical grading relies heavily on physician subjective experience. Moreover, the strong cardiac pulsation artifacts and dynamic filling characteristics of fine collaterals in invasive coronary angiography (ICA) videos make it difficult for conventional static networks to achieve accurate automated assessment. To address these challenges, this paper proposed the first automated Rentrop functional grading framework for ICA videos. First, to mitigate dynamic artifacts and computational costs, we introduce a Background Suppressed Hemodynamic Projection (BSHP) method. This method utilizes physical background subtraction to eliminate skeletal noise and encodes early, mid, and late hemodynamic processes into an RGB color space, achieving lossless compression of high-dimensional videos into 2D representations containing temporal semantics while significantly reducing computational complexity. Second, targeting the microvascular feature extraction difficulty, a Tri-Statistic Edge-Aware (TSEA) Module is designed. By fusing standard deviation pooling and Sobel edge priors, the model's ability to perceive weak collateral textures is significantly enhanced. To validate the effectiveness of the proposed method, extensive experiments were conducted on a self-constructed dataset CCC2026. Experimental results demonstrate that the method achieved 70.93% accuracy, 85.98% specificity, and an AUC of 67.97% in the Rentrop functional binary classification task. The code is available at: https://github.com/yhj0-0/CCCC.

Keywords: Generative AI, Machine Learning, Neural Networks and their Applications
Abstract: Current neuroconnectome generators fall short, textsc{GAN}s ignore topology, wiring models oversimplify biology, and existing autoregressive / diffusion methods fail to capture modular and spatial organization. We propose Neuro-PARSDiff, the first structure-aware autoregressive (AR) diffusion framework for scalable, high-fidelity reconstruction of dense brain connectomes. Existing graph generators treat the connectome as a flat matrix, discarding the modular and hierarchical organization essential for modeling neurodegeneration. Neuro-PARSDiff decomposes each connectome via a permutation-consistent multi-hop structural ranking boldsymbol{psi} into topologically coherent blocks, reconstructing them incrementally via: (i) a learned block-size predictor; (ii) a shared permutation-equivariant diffusion denoiser; and (iii) a hybrid higher-order graph transformer with textsc{GPT}-style causal masking. On textsc{HCP} Connectome (with 1,064 samples, each with up to 1,015 nodes) and textsc{OASIS-3} (756 samples, 1,058 nodes) datasets, Neuro-PARSDiff achieves state-of-the-art fidelity (with textsc{MSE} 0.036downarrow, textsc{CORR};0.82uparrow, textsc{NMI};0.78uparrow, Spectral Distance;0.17downarrow). Augmenting an Alzheimer's textsc{GNN} classifier with our synthetic connectomes raises accuracy by 9.3% and textsc{AUC} by 0.090 over real-data baseline, directly validating clinical utility.

Keywords: Connected and Automated Vehicles, Deep Learning, Intelligent Transportation Systems
Abstract: Deploying accurate object detection for Vulnerable Road User (VRU) safety on edge hardware requires balancing model capacity against computational constraints. Large models achieve high accuracy but degrade sharply under the INT8 quantization commonly used for edge deployment, while small models sacrifice detection performance. This paper presents a knowledge distillation (KD) framework that trains a compact YOLOv8-S student (11.2M parameters) to mimic a YOLOv8-L teacher (43.7M parameters), achieving 3.9x compression while retaining accuracy under INT8 quantization. We evaluate on full-scale BDD100K (70K training images) with TensorRT Post-Training Quantization to INT8. The teacher suffers catastrophic degradation under INT8 (-23% mAP), while the KD student retains accuracy (-5.6% mAP). Analysis shows that KD transfers precision behavior rather than raw detection capacity: the KD student reaches 0.748 precision versus 0.653 for direct training at INT8, a 14.5% relative gain at equivalent recall, and 44% fewer false alarms than the collapsed teacher. At INT8, the KD student exceeds the teacher's FP32 precision (0.748 vs. 0.718) in a model 3.9x smaller, but does not improve recall. For the dataset, detector family, and quantization scheme evaluated here, these findings establish knowledge distillation as an effective approach for deploying accurate, safety-critical VRU detection on edge hardware.

Keywords: Neural Networks and their Applications
Abstract: Automated evaluation remains challenging for mathematical reasoning and code generation, where response quality depends not only on the final answer but also on intermediate reasoning, executable behavior, and task-specific constraints. Static benchmarks often provide only narrow correctness signals, while LLM-as-a-Judge approaches can be sensitive to prompts and may produce weakly supported rationales for tasks requiring precise verification. We introduce Cogito-EVAL, an agentic hybrid evaluation framework that integrates executable verification with rubric-based semantic assessment. Cogito analyzes each evaluation instance, routes it to appropriate tools, collects structured evidence, and aggregates validated errors through dimension-level judgment and severity-aware scoring, yielding calibrated decisions with auditable traces. Experiments on mathematical reasoning and code generation benchmarks show that Cogito achieves stronger agreement with human judgments than representative automatic metrics and LLM-based evaluators. The results indicate that combining tool-grounded evidence with structured multi-stage judgment improves both evaluation accuracy and interpretability, offering a practical bridge between executable checks and semantic assessment.

Keywords: Machine Learning
Abstract: Irrigation scheduling under salt stress is a complex sequential decision-making problem where traditional reinforcement learning (RL) struggles with unstable early exploration and stage-varying physiological sensitivities. To address these bottlenecks, we propose Proximal Policy Imitation Learning (PPIL), a novel framework integrating hierarchical RL and imitation learning. In PPIL, a high-level policy conducts stage-aware long-term planning, while a low-level policy executes precise short-term continuous water allocation. To ensure action safety and sample efficiency, PPIL leverages crop-model-generated expert trajectories via a dynamically decaying Kullback-Leibler (KL) divergence to constrain early exploration. Furthermore, long-term salt accumulation risks are explicitly embedded as safety penalties. Comprehensive experiments in the AquaCrop-RL coupled environment demonstrate that PPIL significantly outperforms classical baselines (e.g., PPO, SAC) in convergence speed, training stability, and robustness against meteorological disturbances.

Keywords: Medical Informatics, Generative AI, Deep Learning
Abstract: Accurate and responsible medical question answering (QA) is important in healthcare, where complex cases require factual knowledge and nuanced reasoning. Existing medical QA systems, typically based on single-agent architectures and static retrieval, often lack adaptability, persistent memory, and structured decision-making. This work introduces an adaptive memory and reflection (AMR) agentic system, a multi-agent framework in which specialized agents use dedicated memory and reflection-based feedback to retrieve relevant prior cases and improve subsequent reasoning. Complexity assessment routes questions through solo, collaborative, or escalated workflows, while consensus and ethical overseer modules support reasoning consolidation and output review. Evaluation on MedQA and MedMCQA demonstrates strong performance compared with several baselines. Ablation studies show that combining agent-specific memory, reflection, and external retrieval yields the strongest performance. These findings highlight the potential of structured memory and feedback for developing more trustworthy medical agents. The source code is publicly available at https://github.com/mm-air/AMR-Agent.

Keywords: Image Processing and Pattern Recognition, Visual Analytics, Communication, AI and Applications
Abstract: UAV-to-satellite geo-localization aims to estimate the geographic position of a UAV image by matching it against a geo-referenced satellite gallery. In cross-dataset deployment, however, a retriever trained on one benchmark often exhibits a ranking bottleneck on unseen datasets: the correct match is frequently contained in the top retrieved candidates, but is not ranked first. To address this issue, we propose a training-free three-stage reranking framework that combines a frozen base retriever, vision-language-model (VLM)-based listwise reranking, and confidence-aware score fusion. Given a UAV query and the top-K satellite candidates returned by the retriever, a VLM jointly compares the candidate set and produces a structured ranking together with a confidence estimate. The resulting listwise signal is then fused with the base retrieval score to obtain the final ranking. Under transfer from University-1652 to SUES-200, evaluated on sampled reranking groups, the proposed framework substantially improves Top-1 accuracy across all four flight altitudes, reaching up to 99.5%. However, experiments on sampled reranking groups constructed from DenseUAV reveal a clear limitation: when geographically adjacent candidates become visually near-indistinguishable, VLM-based reranking provides little benefit and may even amplify ranking errors. These results show both the practical value and the applicability boundary of VLM-assisted reranking for UAV geo-localization.

Keywords: Image Processing and Pattern Recognition, Machine Vision, AIoT
Abstract: Pedestrian tracking in video is important for applications such as surveillance and scene analysis, but it becomes more challenging in crowded scenarios with frequent occlusions and rapid motion. This paper analyzes how detector architecture, model scale, scene density, and inference regime affect multi-object pedestrian tracking. We compare the YOLO11, RT-DETR, and YOLO26 detector families with the ByteTrack and BoT-SORT trackers on sequences from the MOT17, MOT20, and SOMPT22 benchmarks. For YOLO26, we also compare traditional overlap filtering via Non-Maximum Suppression (NMS) with end-to-end inference. The results show that RT-DETR tends to achieve the best tracking quality, especially in denser scenes, albeit at a higher computational cost. In contrast, the medium and large YOLO26 variants offer a more favorable balance between quality and computational cost. Furthermore, removing NMS from YOLO26 did not yield consistent tracking gains. Taken together, these results provide a practical guide for selecting detectors and trackers for multiple-object tracking (MOT) tasks based on application priorities.

Keywords: Networking and Decision-Making, Communications, AI and Applications
Abstract: This paper presents a fault-tolerant coordination framework for multi-UAV networks operating under online expansion (OE) and random faults (RF) conditions. The problem is formulated as a multi-agent reinforcement learning (MARL) task based on the Proximal Policy Optimization (PPO) algorithm. To address dynamic topology variations, we develop a Recovery-shaped Cooperative Graph PPO (ReCoG-PPO) approach that integrates a graph readout for scale invariance, inter-UAV communication for explicit cooperation, and an exponential moving average (EMA)-based recovery shaping to accelerate convergence. The framework enables decentralized execution with varying numbers of agents and under different fault conditions. Simulation results demonstrate that ReCoG-PPO achieves faster recovery and lower interruption than both graph-based and non-graph baselines, while maintaining high throughput, fairness, and per-user rate in fault-tolerant and dynamically scalable scenarios.

Keywords: Robotic Systems, Decision Support Systems, Intelligent Power Grid
Abstract: 子组件，在变电站中是接触丰富的 多阶段操作任务，具有窄净空 校准和阶段依赖控制要求。标准扩散政策通过 单作用分布，这通常会导致不稳定 舞台转换和不准确的局部组装动作。前往 针对这个问题，我们提出了一个阶段条件化方法 明确包含组装的扩散政策 将信息转化为行动生成。方法 支持跨阶段感知控制，涵盖抓取、转移， 校准、插入，以及进一步使用轻量级 可行性筛选步骤以提升执行效率 稳定。基于MuJoCo的双臂模拟器实验 将拟议方法与BC、ACT、扩散政策进行比较， DP3和SmolVLA的演示预算均低于匹配，且 噪音设置。结果显示成功率更高， 动作误差更低，且比 基线，表明显式阶段条件反射为 适用于多级绝缘子组装。

Keywords: Medical Informatics, Transfer Learning, Deep Learning
Abstract: Abstract— To address the challenges of pseudo-label noise and anatomical structure inconsistency in Source-Free Unsupervised Domain Adaptation (SFUDA), this paper proposes a reliability-driven source-free medical image segmentation framework (RDVC) based on two-view consistency. The framework constructs complementary views through style and structural augmentations, and utilizes a dual-voting mechanism to eliminate abnormal predictions, thereby enhancing the robustness of pseudo-labels. Simultaneously, a refinement mechanism based on uncertainty modeling is introduced, which combines predictive entropy maps and certainty regularization for dynamic error correction, driving the model to generate structurally coherent and high-confidence segmentation results.Experiments on public fundus image datasets demonstrate that RDVC suppresses noise interference in self-training, significantly improving cross-domain segmentation performance and robustness in privacy-preserving scenarios. Index Terms— Source-Free Unsupervised Domain Adaptation (SFUDA), medical image segmentation, reliability-driven, two-view consistency, uncertainty modeling.

Keywords: Robotic Systems, Cyber-physical systems, Human-Collaborative Robotics
Abstract: Bimanual mobile manipulation requires tight integration between high-level semantic reasoning and safe, compliant physical interaction. End-to-end models often encode this relationship opaquely, whereas classical controllers lack semantic context. This paper presents GenerativeMPC, a hierarchical cyber-physical framework that maps visual and linguistic context to interpretable control parameters for a bimanual mobile manipulator. A Vision-Language Model with Retrieval-Augmented Generation (VLM-RAG) selects velocity limits and a safety margin for a reduced task-space Whole-Body Model Predictive Controller (MPC), while scheduling virtual stiffness and damping gains for an impedance-admittance controller, enabling context-aware compliance during human-robot interaction. An experience-driven vector database retrieves previously validated parameter sets without model retraining. Human-aware navigation is evaluated in simulation and on a differential-drive platform, where the commanded velocity is reduced by 60% near a detected person while maintaining the selected experimental separation threshold. Bimanual manipulation is evaluated in simulation, achieving an 85% overall task-success rate and final end-effector errors below 5 mm, with the corresponding hardware gains proposed for future deployment.

Keywords: AI and Applications, Deep Learning, Neural Networks and their Applications
Abstract: The development of Financial Natural Language Processing (NLP) systems for Brazilian Portuguese is hindered by the scarcity of domain-specific datasets. To address this, we present a large-scale corpus of approximately 160,000 Brazilian Portuguese financial news articles collected from web sources. To establish initial benchmarks for a weakly supervised multi-class classification approach based on semantic similarity, we annotated subsets of this corpus both manually and using a Large Language Model (LLM). We evaluated several similarity-based methods – including cosine similarity, K-Nearest Neighbors (KNN), and FAISS – using multilingual semantic embeddings. Our experiments reveal that the task is highly sensitive to label quality and class distribution. On the manually labeled subset, KNN achieved the best performance (47.3% accuracy, 31.9% F1-score), closely followed by FAISS. Conversely, on the LLM-annotated subset, while accuracy reached approximately 50%, F1-scores dropped to near 20%, highlighting the negative impact of label imbalance and noisy weak annotations. These findings establish foundational baselines for a challenging task, underscoring the critical need for high-quality manual annotations. Ultimately, this dataset and our preliminary results provide a reference point for future research in Brazilian Portuguese Financial NLP.

Keywords: Human-Centered Transportation, Decision Support Systems, Human-Machine Interaction
Abstract: Public-transit systems are increasingly intelligent at route planning, schedule retrieval, and disruption management, yet they still largely assume that riders already know what a station is good for in a concrete situation. For unfamiliar users, the harder problem often comes before route execution: whether a station should be understood as a gateway, a fast transfer conduit, a low-burden waiting place, or an easy rendezvous point for the task at hand. This paper studies that pre-routing problem as station-level semantic mediation. We organize the study around three questions: what design requirements for unfamiliar-user station sensemaking emerge from interview-informed requirement elicitation, whether semantic Station Sense Cards improve scenario-sensitive understanding relative to factual baseline cards, and whether locally constrained cards improve representational integrity and confidence calibration relative to semantic cards derived only from objective signals. We address these questions with a locally constrained representation that combines objective station signals, structured local semantic constraints, and explicit uncertainty disclosure. Across 24 station-scenario cells, 72 unfamiliar participants completed 1,728 trials and 24 local auditors contributed 576 integrity ratings. The locally constrained condition achieved the highest agreement with local acceptability envelopes for both role judgments (67.5%) and fit judgments (60.4%), outperforming both the factual baseline (46.4%, 44.6%) and the objective-only semantic condition (58.7%, 53.6%). The separation widened in contested cases, where role agreement reached 72.6% under local constraint. Local auditors also rated the locally constrained cards highest on authenticity (5.23/7) and uncertainty adequacy (4.54/7), whereas the objective-only condition elevated confidence (72.9%) without comparable calibration gains. The contribution is a human-centered, uncertainty-aware methodology for station representation that improves unfamiliar-user judgment while preserving local contestability and auditability.

Keywords: Human-Machine Interaction, Generative AI, Human-Computer Interaction
Abstract: Generative AI can make open-ended knowledge work faster and more fluent, but the same workflow can also shift credit, control, and capability from the human to the system. Prior work shows strong productivity and content-generation benefits from AI-assisted writing and creativity support, yet recent HCI and behavioral evidence also points to losses in ownership, accomplishment, and independent confidence when AI occupies too much of the task. A key gap is that most studies either compare AI versus no AI, vary only one workflow dimension at a time, or operationalize “AI-led” collaboration through framing rather than actual process order. We address this gap by jointly studying two workflow variables: help structure (substitution versus scaffold) and initiation order (AI-first versus user-first). We implemented a Wizard-of-Oz prototype and evaluated the full 2×2+1 design in a between-subjects study around an open-ended product-analysis task. Across personal contribution, psychological ownership, self/AI attribution, autonomy frustration, independent self-efficacy, and perceived learning gain, the most adverse profile emerged when AI produced a full answer before the user contributed. In contrast, a user-first scaffolded workflow tracked closely with the no-AI baseline on several agency-relevant outcomes while retaining high perceived learning gain. The paper contributes a workflow-centered account of why AI use can feel empowering in one design and psychologically substitutive in another, a reproducible paradigm for disentangling help structure from sequencing, and an actionable design principle: keep the human upstream, then let AI scaffold downstream.

Keywords: Situation Assessment, Generative AI, Multimedia Computation
Abstract: Situation tracking on tabular data is typically solved with rule-based or Complex Event Processing (CEP) systems, in which the event patterns of interest (the what? to monitor) can be symbolically described once at compile time, only requiring to match the events at runtime against these situation templates to detect matching event constellations. Systems processing non-tabular data like videos or images instead have relied on Deep Learning-based event recognition approaches so far: These inductively learn to generalize to the event patterns of interest from a plethora of labeled examples in their training data, which are labor-intensive to obtain. Given the general-purpose semantic image understanding capabilities of the recently introduced multimodal Foundation Models, we thus ask whether these offer the potential to instantiate CEP on video data: Instead of having to annotate thousands of concrete example instances in video data for training black-box video classification models, we examine if the situation types of interest can only be described once on a symbolic level, converting the problem of situation tracking in video into classical CEP. We thus contribute a training-free neuro-symbolic situation tracking approach to detect situations in video streams, which connects our prior work on a symbolic notation for modeling and tracking evolving situations with sub-symbolic object and relationship detectors to solve the symbol grounding problem for video input. The feasibility of our proposed neuro-symbolic tracking is evaluated on the domains of patient safety monitoring and pedestrian crossing intent tracking.

Keywords: Smart Buildings, Smart Cities and Infrastructures, Cyber-physical systems, Fault Monitoring and Diagnosis
Abstract: While prior work on millimeter-wave (mmWave) radar sensing has focused on human-scale targets, detecting small objects such as mice remains challenging due to their low radar cross section and weak reflections in cluttered indoor environments. We present mMouse, an mmWave radar system for detecting mouse activity as an early indicator of rodent infestations. mMouse models environmental background conditions using matched background recordings and formulates detection as an anomaly detection problem using Deep Support Vector Data Description (SVDD), calibrated with background-only measurements at deployment, enabling operation without labeled mouse-present data. A transformer-based sequence model further estimates mouse trajectories from temporal radar signals. We evaluate mMouse across sensing ranges of 15-125 cm, achieving up to 99% presence-detection accuracy, with trajectory RMSE as low as 11.5cm. Experiments under representative non-line-of-sight indoor conditions, including drywall and cluttered wooden barriers, show how detection performance is governed by sensing geometry and occlusion in addition to distance. These results highlight the potential of mmWave radar as a sensing modality for early rodent detection in indoor environments.

Keywords: Cyber-physical systems, System Architecture, Cyber Systems and Engineering
Abstract: Engineering a cyber-physical or robotic system requires one authoritative model to be propagated, without drift, into a heterogeneous toolchain of simulators, runtime stacks and analysis tools. A common integration practice is a network of pair-wise converters, which scales quadratically in debugging cost, erases n-ary relations, and cannot state cross-view invariants as first-class propositions. HyMeKo replaces that topology with a hub and spokes. The hub is a canonical signed-typed hypergraph intermediate representation (IR) carrying a content-addressable digest; the spokes are a single fixed dispatcher interpreting per-format templates, so that emission-specific knowledge lives in data rather than in code. The pipeline factors into compile, project to sparse tensor, and emit to target-format text, over which four properties follow from the design: alias invariance, content-addressability, projection-emission independence, and bounded storage overhead. Validation uses robotic description generation, a multi-view domain: one model propagates into six views (the Unified Robot Description Format, URDF; the Simulation Description Format, SDF; the MuJoCo XML Format, MJCF; Graphviz DOT; a Gazebo gz sim world; and Mermaid) in under a millisecond for the benchmark fixtures, and the emitted bundle launches under ROS 2 without hand editing. Compilation and projection scale near-linearly to ten million signed incidences, a 68-link humanoid emits all six views in 1.3 ms, and co-emission from one IR preserves a root-link entity and its mass attribute that a pair-wise conversion chain silently drops. The result is a domain-independent substrate for auditable multi-view engineering, with robotics supplying the measurements rather than the assumptions.

Keywords: Medical Informatics, Cyber Systems and Engineering, Information Assurance and Intelligence
Abstract: Rapid sharing of electronic health records (EHRs) across institutions is essential for collaborative healthcare, yet becomes especially critical in emergency scenarios where patient consent cannot be obtained. Existing blockchain-based EHR systems lack an efficient and auditable ``break-glass'' mechanism that enables timely data access while preserving privacy and accountability. To address this, we propose B-SHARE, a blockchain-based framework for secure cross-institutional EHR sharing with collaborative emergency access. B-SHARE adapts the Burmester-Desmedt protocol to enable multiple medical institutions to collaboratively generate re-encryption keys without patient consent, governed by smart-contract-enforced (t,n)-threshold policies with full on-chain auditability. Meanwhile, we design a lightweight proxy re-encryption mechanism that offloads heavyweight cryptographic operations to trusted edge devices through pre-computed differences, reducing the cloud proxy overhead to constant-time O(1). A hybrid on-chain/off-chain architecture with a private IPFS cluster further ensures privacy and storage performance. Experimental evaluations demonstrate that our scheme achieves low emergency access latency and excellent transaction throughput, confirming its suitability for practical healthcare deployment.

Keywords: Communications, Distributed Intelligent Systems, Machine Learning
Abstract: Massive Multiple-Input Multiple-Output is a key technology for enabling high data-rate transmissions in Fifth Generation (5G) networks, where base stations (BS) are equipped with a large number of antenna elements to enhance spectral efficiency, coverage, and throughput. In Frequency Division Duplex (FDD) systems, massive MIMO relies on Channel State Information (CSI) feedback from user equipment (UE), with feedback overhead increasing linearly with the number of antennas. The conventional CSI feedback mecha- nisms standardized by the 3rd Generation Partnership Project (3GPP) are limited by the coarse granularity of the codebook, which can render massive MIMO impractical. In this paper we proposed an autoencoder-based solution to compress CSI at the UE side and reconstruct it at BS. The proposed solution is integrated into the 5G link adaptation mechanism and evaluated not only in terms of CSI compression and reconstruction accuracy, but with respect to its impact on transmission performance. The results demonstrate that the proposed solution accurately reconstructs CSI data, achieving a correlation of 0.99, and outperforms the conventional CSI feedback scheme by approximately 21% in terms of throughput. Error Vector Magnitude (EVM) analysis further indicates enhanced transmission reliability under the proposed method.

Keywords: Digital Twin, Cyber-physical systems
Abstract: Digital twin technology has revolutionized monitoring and simulation in manufacturing, smart infrastructure, and healthcare. To our knowledge, prior digital-twin applications in education have primarily focused on physical learning environments, smart campuses, or instructional simulation. This work shifts the unit of synchronization and intervention to curriculum structure itself. This paper introduces the Curriculum Digital Twin (CDT)—a novel framework that models a university curriculum as a living system of interconnected knowledge components, with a continuously updated cyber-layer mirror reflecting realtime student progression data. We formalize the curriculum as a directed weighted graph in which nodes represent course units and edges encode prerequisite dependencies and knowledge transfer efficiency. A Graph Neural Network (GNN) trained on synthetic longitudinal cohort data, calibrated to publicly available graduation-rate distributions and course-level attrition statistics, predicts dropout-risk hotspots one to two semesters ahead of occurrence using only past cohort data, with no information leakage from the predicted period. A model-based simulation engine enables administrators to run counterfactual curriculum-restructuring experiments before institutional commitment. Experiments on a synthetically generated four-year engineering curriculum demonstrate that the proposed CDT-GNN achieves F1 = 0.851±0.023 on bottleneck detection, significantly outperforming a standard Graph Convolutional Network (GCN) baseline (t = 6.92, p = 0.0023, paired t-test), and that simulated prerequisite-restructuring interventions project improvements of up to +1.7% in four-year graduation rates with a simultaneous −2.4% reduction in the equity gap between high- and low-preparation student subgroups.

Keywords: Cyber Systems and Engineering, Neural Networks and their Applications, Deep Learning
Abstract: The rapid proliferation of low Earth orbit (LEO) mega-constellations has intensified the demand for timely satellite maneuver prediction to support collision avoidance and space traffic management. Existing approaches predominantly treat maneuver prediction as a post-hoc detection task, offering no advance warning capability. This paper proposes PI-MMPF-Lite, a Physics-Informed Multi-scale Maneuver Prediction Framework that reformulates the problem as a multi-horizon advance warning classification task, simultaneously outputting warning probabilities for four lead-time windows of {30, 60, 120, 240} minutes. The framework integrates a strictly causal physical decomposition pipeline to suppress measurement noise, and constructs an 8-dimensional multi-satellite cooperative feature set by fusing the primary satellite's orbital state, neighbor satellite detection outputs, and inter-satellite phase differences, enabling the model to capture the formation maintenance coordination patterns inherent to constellation operations. A lightweight LSTM classifier is trained with an adaptive-horizon Focal BCE loss to handle severe class imbalance. Experimental results demonstrate that PI-MMPF-Lite significantly outperforms XGBoost and EphemDiff on event-level primary metrics, achieving a 90% maneuver onset localization success rate on the test set. The results validate that combining physics-informed priors with multi-satellite cooperative information substantially improves the accuracy and robustness of advance maneuver warning during the satellite parking phase.

Keywords: Energy Systems, Adaptive Systems, Sustainable Systems
Abstract: State of charge (SoC) is a key state variable in battery management systems, while battery aging–induced variations in state of health (SoH) often lead to significant performance degradation of data-driven SoC prediction models. Domain adaptation has been widely explored to mitigate distribution shifts in cross-SoH battery state estimation by aligning feature distributions across different health conditions. However, most existing approaches perform global feature alignment and overlook the temporal inconsistency of battery dynamics during aging, where different time steps may exhibit varying transferability across domains. In this paper, we propose an unsupervised domain adaptation framework for time-series–based SoC estimation. The proposed method employs a recurrent neural network to model multivariate battery signals and extract time-step–level hidden representations. Instead of performing global feature alignment, the distribution discrepancy between source and target features is explicitly computed at each time step. These discrepancies are then aggregated through an adaptive temporal weighting mechanism to construct a temporally structured domain alignment constraint. By integrating this time-weighted alignment with the SoC regression objective, the model learns aging-robust temporal representations that are more transferable across different SoH conditions. Extensive experiments on a public full life-cycle battery dataset demonstrate that the proposed approach consistently reduces SoC prediction errors and achieves more stable performance under battery aging scenarios.

Keywords: Networking and Decision-Making, Adaptive Systems, Sustainable Systems
Abstract: Network Traffic Classification (NTC) is fundamental for network management and security analysis. Despite high accuracy, deep learning models are often over-parameterized, leading to high computational cost and limited deployment efficiency. Moreover, retraining on resource-constrained edge devices is difficult, making fine-tuning-based methods impractical. To address these challenges, we first propose a dual-perspective node importance estimation method for the efficient model. Node importance is evaluated from structural and functional contributions, enabling the direct removal of unimportant nodes while preserving performance. The network is modeled as a layered weighted graph, where structural contribution captures connection strength and functional contribution reflects prediction influence. A unified importance score is constructed to quantify node significance. Moreover, an adaptive model structuring mechanism further adjusts the model under varying traffic loads and computational constraints. The proposed method is evaluated on the ISCX VPN-nonVPN dataset and a self-collected dataset, and the experimental results show that under a fixed accuracy threshold the proposed method achieves the highest complexity reduction among existing approaches without fine-tuning while maintaining competitive performance. In addition, the adaptive mechanism effectively handles dynamic traffic loads and alleviates network congestion, making the method suitable for real-world deployment.

Keywords: Human-Machine Cooperation and Systems, Supervisory Control, Human-Machine Interaction
Abstract: Correcting errors in temporally extended video understanding is costly because existing multimodal pipelines expose little intermediate state for inspection, forcing annotators to revisit raw video and reconstruct temporal semantics from scratch. We present IMPACT-CYCLE, a supervisory multi-agent framework that reformulates long-video understanding as iterative claim-level maintenance of a shared semantic memory comprising typed claims, a claim-dependency graph, and a provenance log. A memory constructor initializes the semantic state, while role-specialized verification agents, operating under explicit authority contracts, examine local grounding, temporal consistency, and global semantic coherence. Their evidence is fused by an arbitration agent, which escalates unresolved claims to a human supervisor and triggers dependency-localized re-verification only within the affected claim neighborhood after accepted edits. In our evaluation, supervisory decisions are simulated by an oracle to estimate the upper bound achievable under accurate human intervention. On 240 evaluation videos, the oracle-assisted setting improves VQA accuracy from 0.71 to 0.79. Dependency-localized re-verification further reduces post-edit model calls per accepted edit from 42.3 to 8.7 (4.8x) relative to full re-verification. Code are publicly available at https://github.com/MKong17/IMPACT_CYCLE.

Keywords: AI and Applications, Neural Networks and their Applications, Computational Intelligence
Abstract: Rain streaks in videos severely degrade visual quality and impair downstream vision tasks. While existing event-based methods exploit high-temporal motion cues, they often overlook frequency-domain information, limiting their ability to distinguish rain from complex dynamic backgrounds. To address this, we propose the Wavelet Spatio-Temporal Network (WSTNet), a multi-domain framework that deeply fuses event streams, video sequences, and wavelet transforms. WSTNet employs a stagewise architecture, progressively refining rain removal via stacked Cascaded Rain Cleaners (CRC). Central to our design is the Structure Aware Module (SAM), powered by Wavelet Spatio-Temporal Attention (WSTA). By integrating wavelet decomposition with attention mechanisms, WSTA explicitly separates high-frequency rain noise from low-frequency structural features in the spectral domain. Extensive experiments on multiple benchmarks demonstrate that WSTNet outperforms state-of-the-art approaches, achieving superior structural preservation, temporal consistency, and robust fine-grained rain removal.

Keywords: Image Processing and Pattern Recognition, Deep Learning
Abstract: Camouflaged object detection (COD) seeks to segment objects that are visually indistinguishable from their surroundings. Existing methods often fall short due to their reliance on implicit feature learning without explicit foreground–background modeling. We present a prompt-driven framework that introduces Multi-Scale Visual Prompts (MSVP) as compact, task-specific representations of foreground and background regions. These prompts attend to backbone features through cross-attention at each scale and are hierarchically refined via a cross-scale gated propagation mechanism that transfers semantic cues from deep to shallow layers. To address the cold-start problem inherent in randomly initialized prompts, we propose a Text-Guided Refinement (TGR) strategy that incorporates pre-trained CLIP text embeddings as semantic anchors, coupled with a linear decay schedule that progressively shifts from text-dominated to vision-dominated learning. Extensive experiments on four benchmarks demonstrate that our method achieves state-of-the-art structural prediction quality (Sm) on all datasets and the best overall performance on CHAMELEON.

Keywords: Machine Learning, Computational Intelligence, Artificial Social Intelligence
Abstract: Existing outlier detection methods often rely on sample-level representations that are vulnerable to noise and unstable local structures. Their performance is also sensitive to manually specified neighborhood sizes, while complementary structural-deviation information is insufficiently exploited. To address these limitations, this paper proposes AMGBOD, an adaptive multi-scale granular-ball method for outlier detection. AMGBOD first constructs compact structural representations through a multi-stage granular-ball generation strategy. It then derives ball-specific low-, medium-, and high-scale neighborhoods from mutual-neighbor stability and similarity transitions, avoiding a single neighborhood size shared by all granular-balls. A dual-branch mechanism subsequently integrates multi-scale local abnormality with sample deviation from mean-shifted granular-ball centers, and weighted rank fusion alleviates the incompatibility between score distributions. Experiments on 20 public datasets show that AMGBOD obtains the best average AUC and average rank among the compared methods, demonstrating competitive performance across diverse data characteristics.

Keywords: Expert and Knowledge-Based Systems, Knowledge Acquisition, Generative AI
Abstract: Large language models (LLMs) have demonstrated strong natural-language understanding, yet their deployment in high-stakes medical settings remains hampered by factual hallucinations and the lack of traceable reasoning. We present HEAR (Hybrid Evidence-Augmented Reasoning), a graph-guided retrieval-augmented framework that alternates between knowledge-graph path exploration and medical text retrieval over multiple reasoning iterations. The relational structure of the knowledge graph steers retrieval toward clinically relevant evidence, while the retrieved text in turn guides and refines subsequent graph traversal. This iterative, bidirectional collaboration grounds LLM reasoning on complementary structured and unstructured knowledge. Systematic comparisons with MindMap and several other baselines on two medical question-answering benchmarks show that HEAR achieves substantial improvements in both answer accuracy and factual consistency. Ablation studies further confirm that the bidirectional coupling between graph reasoning and text retrieval is the primary driver of performance gains, as removing either knowledge source leads to a notable degradation.

Keywords: Systems Safety and Security, Distributed Intelligent Systems, Infrastructure Systems and Services
Abstract: The immutability of blockchain ensures data in-tegrity yet conflicts with legitimate modification requirementsin cross-organizational settings. Existing redactable blockchainschemes predominantly rely on single-authority or threshold-based trapdoor management, which is ill-suited for distributedgovernance where participating entities demand autonomouscontrol and heterogeneous policies. To address these chal-lenges, this paper proposes a multi-authority attribute-basedencryption scheme for controllable cross-domain data mod-ification. First, we construct a multi-authority cross-domainmodification collaborative framework, where each organiza-tion acts as an independent attribute authority managingits own domain keys, achieving decentralized co-governanceof edit permissions. Second, we design a secure multi-partycomputation-based trapdoor joint generation method, enablingauthorities to collaboratively produce the chameleon hashtrapdoor without revealing private key shares. Third, wepropose an attribute-binding mechanism that cryptographicallybinds trapdoor validity to the user’s attribute set and identity,fundamentally blocking permission delegation. Furthermore,a read-write integrated permission control method unifiesdecryption and modification rights within the same attribute-based policy framework. Finally, a smart contract-based trustedcollaboration and attestation mechanism ensures automatedexecution and full auditability. Experimental results show thatcore operations complete within 500 ms, trapdoor generationwith five authorities takes 127 ms, and read-write integrationimproves encryption efficiency by 31%, achieving a practicalbalance between security and performance.

Keywords: Image Processing and Pattern Recognition, Neural Networks and their Applications, Machine Vision
Abstract: Remote sensing oriented object detection is challenged by complex backgrounds, large-scale variations, and arbitrary object orientations. Moreover, long-tailed class distributions further hinder effective learning, where head classes dominate training and limit the performance of tail classes. To address this issue, we propose a Class-Partitioned Dual-Head RCNN (CPDH-RCNN), which incorporates feature refinement for long-tailed oriented object detection. Specifically, classes are divided into head and tail subsets. Two independent detection heads are then assigned to these subsets respectively, alleviating interference between head and tail classes. Additionally, a Feature Refinement Module (FRM) is introduced to enhance multi-scale feature representations and improve proposal quality. The proposed CPDH-RCNN is implemented based on Oriented R-CNN with a PKINet backbone. Experiments on DOTA-v1.0 and DOTA-v1.5 show that the proposed method achieves 80.15% and 73.48% mAP, respectively, with clear improvements on tail classes, and achieves state-of-the-art performance on DOTA-v1.0.

Keywords: Deep Learning
Abstract: Graph neural networks are vulnerable to malicious edge injection attacks, which perturb graph structure and degrade downstream task performance. Effective defense requires both identifying suspicious edges and cleaning the graph without excessively damaging benign connectivity. To address this problem, we propose a boundary-aware cascaded detection and cleaning framework for malicious edge injection attacks. The method first performs supervised suspicious-edge ranking using complementary structural and semantic signals, and then applies boundary-aware partitioning to separate ranked candidates into a primary high-risk region and an uncertain region. A second-stage tail reranker further refines boundary-ambiguous candidates within the uncertain region, enabling more controllable graph cleaning. Experiments on four graph datasets spanning node classification and link prediction show that the proposed framework consistently improves suspicious-edge detection and recovers a substantial portion of the attack-induced performance loss across diverse settings.

Keywords: System Architecture, Adaptive Systems, Computational Intelligence
Abstract: Real-space density functional theory (DFT) applications on heterogeneous supercomputers with explicitly managed local data memory (LDM), such as the Sunway SW26010-Pro, face fundamental coordination challenges: extreme memory constraints transform standard dense matrix-matrix multiplication (DGEMM) kernels into resource coordination problems. This paper proposes a coordination-first execution framework integrating three co-designed mechanisms: state-decoupled exclusive-ownership mapping to eliminate distributed synchronization conflicts, coordinated dual-target fusion to suppress redundant memory traffic through data-locality exploitation, and a lightweight multi-layer perceptron (MLP) surrogate controller for adaptive configuration selection under dynamic self-consistent field (SCF) workloads. Unlike existing BLAS libraries and auto-tuning frameworks, this framework establishes a system-level coordination abstraction specifically addressing the unique requirements of explicit-memory environments.

On the SW26010-Pro, our approach achieves 158.25 GFLOPS per core group (10.5% over swBLAS), a 6.25-fold end-to-end speedup in the ARES DFT system, and 77.5% parallel efficiency across 256 processes, approaching the physical memory-bandwidth limit.

Keywords: Supervisory Control, Human-Machine Cooperation and Systems, Human-Machine Interaction
Abstract: We present IMPACT-HOI, a mixed-initiative framework for annotating egocentric procedural video by constructing structured per-hand event records for Human-Object Interactions(HOI), motivated by their potential use as structured supervision for downstream robot learning from human demonstration. IMPACT-HOI frames this task as the incremental resolution of a partially specified, onset-anchored event state. A trust-calibrated controller selects among direct queries, human-confirmed suggestions, and conservative completions based on empirical annotator behavior and evidence quality. A risk-bounded execution protocol, utilizing atomic rollback, ensures that human-confirmed decisions are preserved against conflicting automated updates. A user study with 9 participants shows a 13.5% reduction in manual annotation actions, a 46.67% event match rate, and zero confirmed-field violations under the studied protocol. The code is publicly available at https://github.com/541741106/IMPACT_HOI.

Keywords: Intelligent Power and Energy Systems, Machine Learning, Adaptive Systems
Abstract: Deep learning models for photovoltaic (PV) power forecasting often face challenges such as black-box opacity, physical law violations, and performance decay caused by concept drift. This research develops an adaptive forecasting framework guided by physics and Shapley additive explanations (SHAP). The proposed framework integrates domain physics with data-driven evolution. We use state-of-the-art (SOTA) forecasting architectures, including Autoformer, FEDformer, and Informer, as the backbone. We embed an asymmetric non-negative penalty into the loss functions to eliminate irrational negative power predictions. For diagnostic transparency, a SHAP-based introspection module monitors the global horizontal irradiance (GHR) contribution ratio to identify feature overfitting. This feedback directs a Bayesian optimization loop via the tree-structured Parzen estimator (TPE) algorithm to align model weights with physical causality. To ensure robustness in non-stationary environments, a lightweight evolution module triggered by SimpleADWIN freezes the SOTA backbone and fine-tunes the terminal projection layer. Validations on the Desert Knowledge Australia Solar Centre (DKASC) dataset demonstrate that the framework rectifies causal biases effectively. The method yields a 17% mean squared error improvement for SOTA models and accelerates online adaptation by 92% compared to global retraining.

Keywords: Reinforcement Learning, Cooperative Systems and Control, Swarm Intelligence
Abstract: In recent years, multi-agent reinforcement learning (MARL) has made significant advancements in the domain of multi-UAV target tracking. However, existing MARL-based methods primarily rely on inter-UAV communication for information sharing, limiting their applicability in communication-denied environments. This paper proposes an Attentive Contrastive Learning-assisted MARL method (ACL) to achieve distributed multi-UAV target tracking in communication-denied environments. Specifically, we introduce a global observation prediction module in which an attention integration module adaptively fuses each UAV’s local observations based on rel- evance, and a teacher-student contrastive framework aligns these fused features to help UAVs perceive latent global states. This enables UAVs to infer missing environmental information, thereby alleviating partial observability issues in distributed execution and enhancing collaboration among UAVs. Experimental results demonstrate that our method outperforms existing baselines and enables real-time autonomous target tracking by multi-UAV in communication-denied environments with obstacles.

Keywords: Generative AI, Systems Safety and Security, Deep Learning
Abstract: 大型语言模型（LLM）的广泛应用 将内容来源作为关键安全保障的优先考虑 问题。而统计水印在 开放域文本，其性能在“低熵”中急剧下降 “陷阱”，比如代码生成和数学推理， 其中高度集中的输出分布会产生信号 消灭。本文提出了熵感知语义 水印检测（EASWD），一个集成的框架 具有动态的局部语义不确定性建模 加权策略而不改变嵌入过程。 通过采用重塑尖峰熵（RSE）和语义学方法 对比残差（SCR），我们构造一个加权 通过S形残余熵检测统计量 权重（SREW）函数。理论分析 证明了SREW近似于最优检测 权重在低熵极限下，提供指数函数 衰变率与假阴性率的界限。实证结果 在MBPP数据集中，表明在1% FPR约束下， EASWD实现了0.642的TPR——相比 基线0.142。此外，F1 分数仍高于 93% 即使ࢷ

Keywords: Image Processing and Pattern Recognition
Abstract: Neural Radiance Fields (NeRF) achieve high-quality 3D reconstruction and novel view synthesis from multi-view images. However, existing methods require clean inputs, whereas real-world data suffers from motion and defocus blur that severely degrade NeRF reconstruction. We present WS-NeRF, a Mamba-driven world-state-aware adaptive deblurring NeRF to jointly handle image degradation and 3D inconsistency. We model NeRF deblurring as a dynamic temporal memory evolution process, and adopt a multi-dimensional world-state observer with MoE mechanism to adaptively calibrate external deblurring prior confidence. Experiments show WS-NeRF improves blurred NeRF reconstruction, outperforms prior methods on PSNR/SSIM/LPIPS, and achieves more stable optimization.

Keywords: Image Processing and Pattern Recognition, Human-Computer Interaction, AI and Applications
Abstract: Dense temporal annotation of procedural activity videos is essential for developing action understanding systems, yet it remains costly because existing tools typically process each human correction as an isolated edit. We present IMPACT-SCRIBE, an interactive framework that turns boundary scribbles into reusable supervision for dense temporal action annotation. The framework combines uncertainty-aware scribble encoding, a lightweight local proposal model, cost-aware query planning, structured dense propagation, and correction-driven adaptation. At each interaction, the annotator reviews a planned boundary query and confirms, adjusts, or rejects the proposed correction. Confirmed edits update the dense sequence labeling, while both accepted and rejected outcomes inform subsequent query selection, effort estimation, and confidence calibration. In a human study with a clip/prelabel initialization imbalance, Full Assist shows descriptively higher strict-tolerance boundary accuracy and lower reported workload than manual VIA, although it does not consistently shorten completion time. Controlled analyses on held-out IMPACT sequences further characterize the roles of the individual components, while 95th- and 99th-percentile system-side latency remains below one second. Together, these results support correction reuse as a practical design principle for interactive temporal annotation. Code and reproduction resources are publicly available at https://github.com/BanzQians/IMPACT_SCRIBE.

Keywords: Machine Learning, Deep Learning, Representation Learning
Abstract: Recent studies have shown that weakly supervised video anomaly detection (WSVAD) methods based on contrastive language-image pre-training (CLIP), by introducing high-level semantic priors, can enhance feature discriminability and thereby improve detection performance. However, these methods often fail to adequately model abrupt temporal variations of anomalies and rely on static textual prompts with limited semantic adaptability, which restricts their ability to capture fine-grained temporal dynamics and evolving semantic cues in complex real-world scenarios. To address these limitations, we present a novel framework, TEA-VAD. We design an SDT module to jointly model local temporal variations and global contextual dependencies. We propose a Temporal Difference-Aware (TDA) learning strategy to learn discriminative temporal representations for improved anomalous boundary localization, and a Dynamic Textual Adaptation (DTA) module to optimize textual representations, thereby improving the model's discriminative capability. Experimental results on UCF-Crime and XD-Violence demonstrate that our method achieves competitive performance in WSVAD, validating the effectiveness of TEA-VAD.

Keywords: Intelligent Transportation Systems
Abstract: Traffic forecasting is a fundamental task in intelligent transportation systems. Recent spatio-temporal models have achieved strong benchmark results, but two issues remain insufficiently addressed: traffic patterns at different nodes can be temporally shifted, and spatial interactions can reflect different structural roles. Many methods still estimate temporal dependencies under shared indices or fixed delay patterns and mix heterogeneous spatial relations in a single learned relation field or fixed fusion scheme. We propose a unified framework that combines Adaptive Temporal Referencing (ATR) and Structure-Guided Relation Modeling (SRM). ATR predicts node-wise temporal offsets and re-parameterizes relative temporal references in attention space, while SRM builds local, topological, and directional relation views and aggregates them through contextual semantic routing. Experiments on four public benchmarks show strong overall performance against representative baselines. Ablation, peak-versus-off-peak evaluation, temporal-offset visualization, and routing observations indicate that ATR improves temporal matching beyond a shared bias and that SRM adapts relation emphasis according to traffic context. The code is available at https://github.com/M78C/ATR-SGRM.

Keywords: Generative AI, Game Theory and Multi-Agent Systems, Machine Learning
Abstract: MAD enhances LLM reasoning capabilities. Debate effectiveness depends on rounds, confidence aggregation, and reasoning effort. We evaluate using GSM8K, MATH, ARC, and MMLU datasets, comparing token consumption and accuracy. Two rounds achieve highest accuracy and token efficiency; accuracy drops from third round. Role differentiation reduces token consumption by 9.4% while improving performance by 16.46%. Our dynamic scheme improves accuracy by 6.79% and reduces token consumption by 68.55% across 26 datasets. This reduction uses only two-round debates.

Keywords: Medical Informatics
Abstract: Clinical diagnosis is an inherently multimodal reasoning task that requires the joint interpretation of medical images, free-text clinical notes, and physiological time-series, while grounding every inference in established medical knowledge. Existing approaches either fuse modalities through end-to-end learned representations that lack transparency or augment large language models (LLMs) with static knowledge bases that cannot adapt to individual patient contexts. In this paper, we propose textbf{MAGRA}, a textbf{M}ulti-textbf{A}gent framework with textbf{G}raph textbf{R}etrieval-textbf{A}ugmented reasoning for medical diagnosis. MAGRA routes heterogeneous patient data to modality-specific specialist agents. A graph-construction agent then assembles their structured outputs into a patient-specific knowledge graph through a two-stage entity-anchored extraction pipeline with strict referential validation. A coordinator agent derives weighted query terms from the patient graph, retrieves corroborating evidence from an external biomedical knowledge graph, and fuses the top-ranked triples back through type-aligned merging with provenance tagging. Finally, a prediction agent reasons over the enriched graph summary, multimodal summaries, and raw data to produce diagnoses with full attribution tracing. On the MIMIC-IV ICD-10 multi-label disease diagnosis task, MAGRA achieves a 32.9% relative F1 score improvement over the strongest baseline, validating the effectiveness of integrating dynamic patient graph construction with retrieval-augmented external knowledge verification for interpretable clinical decision support.

Keywords: Neural Networks and their Applications, AI and Applications, Deep Learning
Abstract: Precipitation nowcasting remains a critical challenge in meteorological forecasting. Existing deep learningbased methods often exhibit issues such as blurred precipitation fields, weakened heavy precipitation echo prediction, and inadequate modeling of multi-scale spatiotemporal dependencies when forecast lead times increase. To address these issues, we propose MSPANet, an enhanced spatio-temporal prediction framework based on the SimVP architecture. MSPANet employs a multi-scale feature fusion module with window-based cross-attention and frequency-domain enhancement, effectively integrating features across spatial scales while preserving coherent precipitation structures. This design alleviates the semantic gap between shallow and deep features, significantly improving the forecasting accuracy of precipitation structures and the preservation of heavy precipitation echoes. In the decoder, we design a temporal convolutional module with compressed attention to enhance the modeling capability of key temporal features in heavy precipitation events. Additionally, we propose a novel loss function that integrates time-weighted multiscale MSE with structural similarity (SSIM) to achieve multiscale structural preservation and intensity-aware optimization. Extensive experiments on two publicly available radar datasets demonstrate the effectiveness and superiority of MSPANet compared to state-of-the-art methods, particularly in heavy precipitation forecasting and in maintaining structural fidelity over longer forecast horizons.

Keywords: Affective Computing, Brain-Computer Interfaces
Abstract: Decoding affective states from heterogeneous physiological signals is challenged by modality misalignment and signal non-stationarity. Drawing inspiration from human cognitive processing, we introduce ReEntE, a novel framework for the synergistic integration of EEG and eye movement data. At its core is a Neuro-Behavioral Alignment module that explicitly bridges the semantic gap between neural signals and behavioral cues, facilitating robust cross-modal synchronization. Furthermore, the core Re-Entrant Expert Network, which integrates an Emotion State Space Model, explicitly simulates neural re-entry to mitigate non-stationarity by capturing dynamic emotional evolution through iterative, expert-driven feature refinement. Complemented by dynamic modality-gated fusion, ReEntE achieves state-of-the-art cross-subject accuracy on the SEED and SEED-IV datasets, validating the integration of neuro-cognitive mechanisms into generalized multimodal affective computing.

Keywords: Cooperative Systems and Control
Abstract: This paper presents a distributed adaptive non-singular terminal sliding mode (NTSM) control strategy for leader–follower voltage tracking in islanded microgrids. The focus is on developing a system that is practical and effective in real-world settings, not just in theory. The main concept is to create a consensus-based adaptive controller compatible with nonlinear distributed generator (DG) dynamics. Rather than just synchronizing with each other, all follower DGs track the reference signal from a single leader. The chosen sliding surface guarantees finite-time convergence to the desired state without encountering singularity issues. An adaptive mechanism is incorporated to manage uncertainties and disturbances, provided they remain within certain limits. Through a Lyapunov-based analysis, it is proven that both voltage and frequency track the reference in finite time, even when DG communication is modeled by a directed graph containing only a spanning tree. In practice, implementation introduces challenges like high computational demands, parameter drift, and the well-known chattering associated with sliding mode control. To address these, a globally coordinated adaptive layer is introduced. This involves a unified adaptive gain (featuring sigma-modification and projection) to stabilize parameters, a smooth switching function to minimize chattering, and a distributed integral restoration term to ensure robust voltage regulation during load changes. These enhancements increase system robustness, scalability, and make the approach suitable for real-time operation in converter-based microgrids. The proposed method was validated on a five-DG islanded microgrid under varying load conditions and communication interruptions. Results demonstrated faster convergence, improved disturbance rejection, and reduced steady-state errors compared to conventional consensus-based secondary controllers. In summary, this adaptive NTSM approach is not only theoretically sound but also practically effective for microgrid voltage restoration.

Keywords: Big Data Computing, Application of Artificial Intelligence, Intelligent Internet Systems
Abstract: Decentralized Finance (DeFi) operates on public blockchains where transactions become irreversible after confirmation, making it necessary to evaluate trust and security risks before transaction finality and under tight latency and decentralization constraints. Existing solutions generally fall short of meeting three goals at once: high detection accuracy, low on-chain gas overhead, and decentralized verifiability. We present TrustManager, a hybrid AI-oracle trust management framework that combines off-chain, interpretable AI analysis with verifiable on-chain enforcement. Off-chain, TrustManager uses a graph-enhanced random forest model with weighted structural features to characterize complex transaction behaviors. On-chain, it employs threshold signatures together with an optimistic challenge mechanism, allowing incorrect oracle outputs to be disputed and penalized through slashing. Experiments on a pooled multi-chain dataset of 560{,}934 labeled transactions (515{,}934 benign from the Etherscan V2 API and 45{,}000 labeled malicious samples from publicly reported exploit incidents) collected from Ethereum, Polygon, and Binance Smart Chain (BSC) show that TrustManager reaches an F1-score of 98.30%. For on-chain enforcement, verifying a single aggregated threshold signature reduces verification cost by 92% compared with verifying 10 individual ECDSA signatures. Our code is public at textcolor{black}{url{https://github.com/nanyun818/TrustManager-Artifact}}

Keywords: Robotic Systems, Adaptive Systems
Abstract: This paper presents a distributed finite-time consensus protocol for a group of cloud-connected Euler-Lagrange systems in the presence of uncertainty. The consensus protocol is designed using nonlinear terminal sliding-mode-based consensus surface vectors combining position and velocity error vectors. The consensus error states use the position and velocity signals of the local and neighboring Lagrange systems. A robust adaptive learning algorithm is used to learn and compensate for bounded uncertainty associated with parameters and other external disturbance uncertainty. Nonlinear terminal sliding mode and Lyapunov theory are used to show the convergence analysis of the proposed protocol. The convergence analysis uses Laplacian matrices and Graph theory to show that the states of a group of nonlinear follower Lagrange systems can reach an agreement by sharing their states to track the states of the leader Lagrange system in finite time. The proof of the convergence has three parts. First, the analysis shows that the consensus error states are continuous, and the parameters are continuous and bounded by compact sets. Second, it shows that the sliding mode motion occurs for each Lagrange system in finite time. Third, convergence analysis shows the finite-time consensus tracking of the position and velocity consensus error states. The design and analysis do not use the exact bounds of the uncertainty associated with the follower Euler-Lagrange systems. The nonlinear terminal sliding mode-based finite-time consensus protocol can provide faster and more robust convergence against uncertainty as opposed to the asymptotic consensus protocol. To demonstrate the convergence, the proposed design is tested on three connected Euler-Lagrange systems.

Keywords: Medical Informatics, Image Processing and Pattern Recognition, Deep Learning
Abstract: ，半监督式医学图像分割缓解了 利用无标签数据来减轻注释负担。然而， 现有的一致性学习方法通常用于处理 不同频率成分均匀分布，忽略 低频结构之间的可靠性差异 信息和高频边界细节。此外，基于当前情况的不确定性估计 预测难以捕捉动态稳定性 培训过程。为解决这个问题，我们提出 频率自适应一致性（FAC）框架 将师生预测差异分解为 低频结构组件与高频 详细分量，并自适应地加权一致性 被强加在他们身上的限制。具体来说，是高斯分布 过滤用于分解师生关系 低频结构组件中的一致性信号 以及高频细节分量稳定一致性 对低频成分施加约束， 而置信归一化熵门（CNEG） 机制与训练轨迹不确定性结合 （TTU）以自适应调制高频一致性A

Keywords: Transfer Learning, Deep Learning, Machine Learning
Abstract: Public scoliosis landmark benchmarks are dominated by preoperative radiographs, yet real-world deployment often involves postoperative follow-up images with substantially different appearance distributions. A common solution is target-aware adaptation, which requires target-domain images, statistics, or optimization signals. This paper explores a stricter and clinically cleaner setting: source-only vertebral landmark detection, where the model never sees the target domain during training or model selection. We propose a style-structure framework built on a heatmap-based vertebral detector. On the style side, random MixStyle is inserted into the early stages of a ResNet34 backbone, mixing source-instance feature statistics during training to simulate appearance variations without target exposure. On the structure side, we introduce Slot-Wise Structure Prior Decoding (SWSPD), which uses a slot-wise vertical position prior to regularize anatomy-aware decoding and promote vertically coherent vertebral ordering. We adopt a public-to-private evaluation protocol: training on a public source benchmark, validating on the same benchmark for compatibility, and testing directly on a held-out private clinical cohort to assess generalization under an unseen implant-induced domain shift. Extensive experiments on both the public benchmark and a private clinical cohort show that the proposed method preserves competitive in-domain performance while improving cross-domain robustness under the strict source-only setting. On the implant-bearing subset of the private cohort, it reduces SMAPE from 18.99 to 13.69 and CMAE from 4.85 to 3.80 compared with the strongest competing source-only baseline.

Keywords: Adaptive Systems
Abstract: For quadrotor UAVs that are susceptible to input saturation and actuator failures, this research suggests an event-triggered finite-time trajectory tracking control approach with active fault tolerance. A coupling model concerning saturation functions, health parameters, and residual disturbances is used to establish the quadcopter UAV's dynamics. The system's finite-time convergence is guaranteed by the nonsingular terminal sliding mode surface, and the linked fault model is handled by a Nussbaum-function adaptive gain. Additionally, radial basis function neural networks (RBFNN) are used to compensate for the threshold, and an adaptive dynamic event triggering mechanism is suggested to lower the system's communication frequency. By demonstrating the closed-loop system's uniform finite-time stability, the Lyapunov stability theory greatly lowers the computational load and communication overhead. Comparative simulations show that, under the linked fault model, the suggested approach successfully avoids saturation boundaries, demonstrating improved fault tolerance and anti-saturation performance while significantly lowering triggering events without sacrificing control quality.

Keywords: Mechatronics, Machine Learning
Abstract: Accurately tracking trajectories with multirotor UAVs remains challenging, primarily because manual tuning of PID gains is not effective under varying conditions and changing mission requirements. This paper introduces a four-phase, machine learning-powered PID control approach for quadrotor UAVs. Phase 1 uses a Ridge Regression model to iteratively find the best outer-loop proportional (Kp) and derivative (Kd) gains. It learns from flight telemetry during a 60-second S-curve helix mission and brings the mean position RMS error down to 0.0607 m. In Phase 2, Fast Fourier Transform (FFT) spectral checks confirm the system’s energy is concentrated below 1 Hz and shows no high-frequency oscillations. Phase 3 puts these gains to the test on a more complex trajectory—a tightening figure-eight (lemniscate) with gyroscopic coupling. The same gains, with no additional tuning, still achieve a 0.097 m position RMS (R^2 = 0.1517, peak thrust 14.01 N), proving their generalizability. In Phase 4, the framework compares seven machine learning regressors as gain optimizers: Ridge Regression, Decision Tree, K-Nearest Neighbours (KNN), Gradient Boosting, AdaBoost, Support Vector Regression (SVR), and XGBoost. These models are dropped into the same iterative control loop for benchmarking. Ridge Regression tops the list, reaching 0.960 on the normalized accuracy metric (RMS ≈0.042 m). In contrast, SVR performs poorly (accuracy 0.159) because it can’t ensure physically valid gain values without enforcing positivity. The other ensemble methods (Decision Tree, KNN, Gradient Boosting, AdaBoost, XGBoost) land close together, each with accuracy between 0.921 and 0.923, supporting the robustness of a gain-from-data approach. These results make a solid case for Ridge Regression as the preferred gain estimator for embedded UAV flight controllers and set the foundation with the first systematic multi-algorithm benchmark for ML-driven PID gain discovery in quadrotor UAVs.

Keywords: Transfer Learning, Generative AI, Deep Learning
Abstract: Zero-Shot Stance Detection (ZSSD) aims to predict stance toward targets unseen during training. A central challenge is to learn stance evidence that remains valid under target shift. Recent contrastive and LLM-enhanced methods have improved ZSSD, but they often do not clearly distinguish target-invariant evidence from target-specific lexical cues, and retrieved knowledge is commonly incorporated as flat text rather than structured transferable reasoning. To address these limitations, we propose KR-DV, a knowledge- and rationale-enhanced dual-view framework. In the text view, KR-DV uses target masking and hierarchical contrastive learning to separate target-invariant and target-specific signals, thereby reducing shortcut reliance under target shift. In the knowledge-graph view, KR-DV retrieves background knowledge, generates neutral rationales, organizes Text, Target, Background, and Rationale as a compact heterogeneous graph, and applies supervised graph contrastive learning to model structured cross-target reasoning patterns. The two views are fused for final prediction. Experiments on SEM16 and VAST show consistent improvements, especially on VAST and in low-resource settings, suggesting the value of combining transferable evidence learning with structured reasoning.

Keywords: Intelligent Transportation Systems, Decision Support Systems
Abstract: The fleet assignment problem (FAP) selects an aircraft type for each flight, thereby determining flight operating costs and maximum passenger capacity. It is a critical step in airline schedule planning. A reasonable fleet assignment achieves precise matching between passenger demand and capacity resources, thereby reducing operating costs. The FAP is NP-hard and subject to constraints including flight coverage and flight connection. This paper proposes a genetic algorithm (GA) framework that integrates connection network (CN) initialization and adaptive mutation operator selection. First, the CN combined with a multiplier penalty method is used to generate high-quality elite solutions. Next, multiple destroy-and-repair operators are introduced to enhance the exploration capability of the GA during the mutation phase. Finally, an adaptive operator selection mechanism based on Thompson Sampling (TS) is constructed. Considering the characteristics of the FAP, strategies such as learning signal decoupling and a fixed β parameter are designed to achieve dynamic adaptation of mutation operators across different search stages. The proposed algorithm is applied to the fleet assignment of all flights operated by an airline on a single day, and the performance of different algorithms is compared. Experimental results show that the proposed algorithm assigns more appropriate aircraft types to flights, reducing operating costs by 1.1%.

Keywords: Systems Safety and Security, Distributed Intelligent Systems, Machine Learning
Abstract: 摘要——联邦学习（FL）极易受到后门攻击，因为服务器对本地数据和训练行为的可见性有限。在非独立且同分布（非IID）环境下，这一挑战更加严重，良性客户端更新的异质性增加了方向、幅度和局部参数行为的变异性。因此，恶意更新更容易融入可信集，而良性但高度异构的更新更可能被错误排除。因此，现有的稳健聚合和异常检测方法在非IID型 FL 中常面临双重挑战：残留的恶意扰动可能留在可信集中，而有价值的良性贡献可能被过度过滤，从而降低了鲁棒性和主要任务性能。为解决这一问题，我们提出了FedMARC，一个多阶段的自适应参考修正框架，用于非IID FL的后门防御。FedMARC由三个协调阶段组成：可信参考构建、关键子空间的差异化特征校正

Keywords: Control of Uncertain Systems, Energy Systems
Abstract: In this paper, we introduce decentralized secondary control for inverter-based islanded AC microgrids with uncertainty. We first develop a decentralized robust adaptive secondary voltage--reactive power sharing control algorithm for islanded microgrids. Lyapunov analysis shows that the controller can asymptotically restore both voltage and reactive power to their respective reference values for all distributed generator (DG) units. Then, we introduce a robust adaptive secondary frequency--active power sharing control algorithm. The Lyapunov method is used to show that the proposed design can restore frequency to the reference value while ensuring accurate power sharing for all distributed generators. The proposed secondary controller does not require information about neighbouring DG units, as the controller is implemented locally using only local information. The design does not require a priori knowledge of the nonlinear dynamical model and uncertainty associated with the DG unit. The simulation results demonstrate on a 3-DG networked microgrid system with uncertainty.

Keywords: Autonomous Vehicle, Machine Vision
Abstract: 4D radar-camera fusion has emerged as a promising sensing paradigm for reliable 3D object detection. Recent methods have considered radar-camera discrepancy through increasingly adaptive fusion designs. However, when modality differences are mainly addressed within the fusion operation, inconsistent local responses may still be aggregated into object features and degrade downstream detection. To address this issue, we propose CAFusion, a consistency-aware 4D radar-camera fusion framework for reliable 3D object detection. CAFusion incorporates consistency-aware interaction at both the pre-fusion and proposal-refinement stages through two complementary modules, Staged Consistency Fusion (SCF) and Proposal-Aware Consistency Enhancement (PACE). SCF first builds a more reliable BEV fusion basis by mitigating global radar-camera differences and refining local feature responses, while PACE further improves proposal-level features by using radar and fused responses to select informative visual evidence around proposal centers. Extensive experiments on View-of-Delft (VoD) and TJ4DRadSet demonstrate the effectiveness of CAFusion, achieving 60.99 mAP and 40.14 mAP, respectively.

Keywords: Conflict Resolution, Cooperative Systems and Control
Abstract: Strategic conflict management in dense low-altitude airspace requires improving safety without incurring excessive computation or unnecessary global trajectory perturbation. This paper proposes a flight risk field (FRF)-aware local replanning framework for unmanned aerial vehicle conflict management in discretized low-altitude airspace. The method retains a standard 4-D conflict definition, abstracts detected interactions into a conflict graph, and decomposes the graph into connected components so that only conflict-relevant subsets are replanned. Within each selected component, an FRF-aware objective jointly considers added time, risk exposure, and path deviation, while conflict-free flights preserve their original intent. In a deterministic simulation testbed with traffic scales of 50, 100, 150, and a 200-flight stress case, the proposed framework improves both tractability and locality. In the 100-flight scenario, FRF-aware local replanning reduces runtime from 20.7 s to 13.9 s and lowers cumulative risk exposure from 42.8 to 34.7, while keeping the increase in average added time modest. These results support an efficient approach to bounded-scope conflict mitigation and risk reduction rather than a strict guarantee of globally zero-conflict traffic plans.

Keywords: Robotic Systems, Modeling of Autonomous Systems, Deep Learning
Abstract: Reliable action generation for multi-stage robotic manipulation remains challenging for Vision-Language-Action (VLA) models. While existing flow-matching VLA policies offer strong multimodal grounding and generalization, they typically employ a single shared action expert, limiting their ability to capture phase-specific control patterns across distinct execution stages. We propose a plug-and-play Phase-Aware Mixture-of-Experts Action Module (PAMAE), as a step towards more reliable phase-consistent action generation. PAMAE replaces the original flow-matching action expert with a sparse expert mixture while preserving the pretrained VLA backbone. PAMAE introduces a phase-aware router that leverages execution-phase cues to allocate action generation across experts, supported by a lightweight phase prediction head and a routing alignment objective. To stabilize specialization, we adopt a two-stage training scheme that first warms up the expert module under the standard flow-matching loss and then optimizes phase-consistent routing under auxiliary supervision. On multi-stage manipulation simulation tasks, PAMAE improves task success by up to textbf{9.2%} over strong VLA baselines. Further ablations show that both phase-supervised routing and staged optimization are essential for the observed gains. Our results highlight phase-consistent expert allocation as an effective mechanism for improving the reliability and action quality of flow-matching VLA policies.

Keywords: Generative AI, Assurance, Information Assurance and Intelligence
Abstract: Foundation models store substantial parametric knowledge acquired during pretraining, yet several questions remain open: how should we evaluate what a model knows, how reliably can it ground new information, and how does it behave when fresh evidence conflicts with prior internal knowledge? Existing benchmarks remain valuable, but most are static and therefore only weakly probe knowledge freshness, conflict resolution, and article-grounded reasoning.

We introduce emph{LiveNews-MCQ}, a continuous evaluation framework that uses live news as a dynamic knowledge source for measuring knowledge acquisition, representation, verification, and application in foundation models. For each fresh article, multiple models generate grounded multiple-choice questions (MCQs) with answer keys, and all models then answer the pooled question bank. This cross-model protocol jointly evaluates question-construction fidelity and answer accuracy on time-stamped, externally verifiable content.

Across 20 live-news articles, 9 model variants, 143 generation records, and 15 answer-stage records, the framework reveals large differences in how models interpret newly observed information, maintain factual consistency, and follow structured output requirements. The strongest model in the available runs, texttt{gpt-oss:120b}, leads both generation fidelity (mean texttt{cognitive_correctness} =0.930) and weighted answer accuracy (69/72=95.8%). LiveNews-MCQ therefore complements static benchmarks by focusing on the textbf{knowledge lifecycle}: acquisition from fresh evidence, representation through question generation, verification through cross-model answering, and grounded application in reasoning.

Keywords: Multimedia Computation, Generative AI, Interactive and Digital Media
Abstract: Audio-driven talking-head video synthesis remains a central challenge in the field of digital human generation. Despite notable improvements in audio-driven 2D facial animation in terms of visual fidelity and lip synchronization, existing editing methods still face limitations in fine-grained control of facial motion, interpretability of expression editing, cross-subject generalization, and temporal consistency. To address these issues, we present MorphTalk, a 2D talking head synthesis framework that utilizes a 3D mesh as an intermediate representation. The system maps speech features to time-aligned facial mesh sequences, achieves localized and precise modulation of expressions through an interpretable BlendShape-based mesh editing module, and uses a high-performance renderer to synthesize portrait videos with high fidelity and temporal consistency. This framework enhances the precision and interpretability of expression editing, improving generalization across different views and motion conditions. Extensive experiments across diverse subjects demonstrate significant improvements in expression controllability and visual quality over the state of the art.

Keywords: Information Assurance and Intelligence, Big Data Computing, Systems Safety and Security
Abstract: Phishing detection remains locked in an arms race over mutable features that attackers continuously change, like URLs, DOM structures and visual appearance, forcing existing detection methods to update continuously. However, attackers aiming to steal credentials must follow an invariant operational core: presenting a form, capturing inputs, and exfiltrating data. In this paper, we exploit this attacker-constrained invariant to build a detection system that resists the evasion tactics defeating existing classifiers. Starting from a threat model that separates mutable surfaces from immutable credential-collection logic, we formalize the Phishing Path Credential Indicator (PPCI) property and verify with Z3 that a PPCI-aware graph pruner preserves every credential-reachable path. Empirical validation on 4,163 real-world traces confirms that this interaction property sharply discriminates phishing from benign pages, a difference of 13.2% in absolute terms (i.e., 84.7% vs. 71.5%). Guided by these results, we design PHISHTRACE: a deep-interaction crawler forces latent attack logic to execute, an Enhanced Interaction Trace Graph (EITG) encodes the resulting behavioral trace, and a Stacked Multi-View Ensemble (SMVE) classifies 125 features across 6 complementary views. In deployment-style evaluation on all 4,163 curated traces, PHISHTRACE-SMVE reaches F1=0.9598 and AUC=0.9924. On the matched 1,701-trace ITG-valid subset used for same-data graph comparisons, the PHISHTRACE family remains ahead of graph-only baselines. Under the implemented adversarial strategies, the multi-view architecture remains highly resilient.

Keywords: Machine Learning, Transfer Learning, Generative AI
Abstract: Single-teacher knowledge distillation enables the transfer of large language model capabilities to smaller models, supporting local deployment and low-cost inference for Verilog generation from natural-language specifications. However, in generative hardware code tasks, teacher outputs are not equally reliable, and blind imitation can cause the student to inherit erroneous or weakly verified patterns, thereby harming compilability and functional correctness. To address this issue, we propose a verification-aware single-teacher distillation framework for hardware code generation. For each input specification, the teacher samples multiple candidate Verilog implementations, which are compiled and simulated to obtain a continuous verification-aware score. This score plays two roles in training: it weights sequence-level knowledge distillation (seqKD) so that fully verified and partially correct candidates contribute with different strengths, and it induces group-wise ranking constraints that provide negative feedback among candidates generated for the same specification. In addition, we design a graded utilization strategy for partially correct samples, preserving useful intermediate supervision through soft-gated seqKD while routing low-quality candidates to ranking-only supervision. Experiments on VerilogEval and RTLLM show consistent improvements in both Compile@1 and SimPass@1 under the same teacher-student setting.

Keywords: Neural Networks and their Applications, Image Processing and Pattern Recognition
Abstract: Camouflaged Object Detection (COD) remains challenging because camouflaged targets often exhibit ambiguous boundaries, inconsistent cross-scale representations, and weak foreground-background discrimination. Although existing methods have explored boundary modeling, multiscale fusion, and complementary cue learning, the interactions among these factors remain insufficiently exploited. To address this issue, we propose the Dual-domain Attention Network (D2ANet), a progressive framework that explicitly constructs and propagates boundary-aware priors for subsequent adjacent-scale interaction and dual-domain refinement. Specifically, the Hierarchical Boundary Extractor (HBE) combines shallow gradient responses with high level semantics to derive boundary-aware priors. Guided by these priors, the Edge-guided Multi-scale Interaction (EGMI) module performs prior-guided adjacent-scale interaction to suppress pseudo-boundary interference and improve cross scale consistency. The Dual-domain Attention (DDA) module further refines the enhanced features by exploiting complementary spatial and frequency cues, thereby improving structural consistency and foreground-background discrimination. Experiments on three benchmark COD datasets demonstrate that D2ANet achieves competitive performance compared with recent representative methods. Extensive ablation, sensitivity, and complexity analyses further validate the effectiveness and design choices of the proposed framework. Code is available at: https://github.com/l08271236-byte/D2ANet.

Keywords: Information Assurance and Intelligence, Generative AI, Expert and Knowledge-Based Systems
Abstract: In the face of increasingly complex and stealthy cyberattacks, traditional Security Orchestration, Automation, and Response (SOAR) systems suffer from delayed responses and limited contextual reasoning, making them inadequate for unknown and multi-stage threats. While large language models (LLMs) offer new opportunities for security automation, their application is hindered by hallucinations, limited domain knowledge, and weak structured reasoning. To address these challenges, this paper proposes an Intelligent Security Playbook Generation System(ISPSYS). ISPSYS integrates three components: a structured knowledge base built on reasoning-oriented Security Atoms that encapsulate prerequisite–action–outcome chains, a dynamic multi-layer context constructor for query-aware retrieval and fusion, and a domain-adapted LLM for security decision-making. Experimental results show that ISPSYS consistently outperforms baseline methods in ROUGE, BERTScore, and METEOR. GPT-4–based evaluation further demonstrates improvements in technical accuracy, step completeness, operational actionability, and logical coherence. These results highlight a practical pathway for transforming general-purpose LLMs into reliable systems for intelligent security operations.

Keywords: Deep Learning, Neural Networks and their Applications, AI and Applications
Abstract: Medical image classification is critical for early disease detection. However, hybrid CNN-Transformer architectures often suffer from semantic drift—the progressive degradation of fine-grained anatomical details caused by over-reliance on high-level semantics. This phenomenon leads to substantial diagnostic errors at low-contrast lesion boundaries, with misdiagnosis rates reaching 10–25% in existing hybrid models. To address this, we propose the Hierarchical Semantic Guidance Framework (HSF-Net), a dual-branch architecture that adaptively refines multi-scale features while explicitly constraining semantic drift. HSF-Net integrates three synergistic components: a Local Feature Enhancement(LFE) module that preserves high-frequency edge details via learnable combinations of four directional edge-filtering kernels; a Coordinate Feature Fusion mechanism that adaptively balances local and global representations through channel-wise attention; and a dual-path Semantic Guidance Layer that injects cross-layer boundary cues and hierarchical multi-scale context into the fusion process. Extensive experiments on six medical imaging benchmarks demonstrate the effectiveness of our approach, achieving a 4.1% accuracy improvement over state-of-the-art hybrid models on COVID-19 CT and a 3.6% gain on Kvasir endoscopic images. Our code is available at: https://github.com/ooolj/HSF-Net.

Keywords: Machine Learning, System Modeling and Control
Abstract: Many physical processes in scientific and engineering computations switch regimes or vary nonstationarily across wide parameter domains. A single global formula can therefore struggle to remain accurate, interpretable, and practical for engineering use. We propose an Error-Manifold Guided Symbolic-Skeleton Mixture-of-Experts (EM-SMoE) framework for this setting. Symbolic regression first extracts a shared analytical skeleton that captures the dominant trend. The residual field of this skeleton then guides adaptive state-space partitioning. Each local expert retains the same skeleton and calibrates only its coefficients, while a gating mechanism blends expert predictions smoothly. This design avoids the boundary discontinuities and inconsistent submodel forms associated with predefined partitions. It also produces stable, explicit expressions that can be audited and embedded directly in solvers. We evaluate EM-SMoE by calibrating the parameter U in a two-temperature model of high-temperature thermochemical nonequilibrium using high-fidelity state-to-state data. The results show that EM-SMoE substantially reduces errors over a wide temperature range while preserving a compact analytical form and approaching the high-fidelity reference.

Keywords: Deep Learning, Application of Artificial Intelligence, Cyber Systems and Engineering
Abstract: ，网络入侵检测在 保护着现代网络基础设施，但它依然存在 由于其规模、异质性增加，以及 网络流量的失衡以及快速流量 攻击行为的演变。现有入侵检测 这些方法通常依赖于监督学习， 手工制作的特征，限制了它们的泛化 能够进行看不见的攻击。近期基于图和 自我监督方法通过利用这些方法展现出潜力 网络拓扑与基于重建的学习; 然而，大多数模型建模的是局部结构 依赖关系或全局时间模式单独存在， 采用简化的聚合策略，忽视这些 细粒度流动相互作用，且受限于有限的 单阶段重建导致的异常辨别 机制。我们提出了E²-GIDS，一种新的基于图的方案 入侵检测框架，联合捕获本地 结构依赖与全局时间共现 网络流量中的模式。E²-GIDS引入了 边缘

Keywords: Deep Learning, Machine Learning, Heuristic Algorithms
Abstract: Few-shot class-incremental learning (FSCIL) in medical image classification faces the dual challenge of long-tailed base-class distributions and limited annotations for novel categories. We present a unified framework that addresses these two issues in a coordinated manner. An uncertainty-guided sampling module is first introduced to enrich informative tail samples and alleviate bias in the base stage. A margin-modulated learning strategy is then employed to preserve generalized representations while improving class discrimination during incremental adaptation. In addition, auxiliary pseudo-labeled data are used to further expand the representation space before novel classes are introduced. Experiments on two MedMNIST benchmarks under both domain-shifting and non-domain-shifting settings show that the proposed framework consistently improves average accuracy and reduces performance degradation compared with strong baselines. These results demonstrate the effectiveness and practicality of the framework for realistic medical FSCIL scenarios.

Keywords: Distributed Intelligent Systems, AIoT, Cyber Systems and Engineering
Abstract: Mobile edge computing (MEC) supports latency-sensitive tasks by deploying computing resources at the network edge. However, individual edge nodes are resource-limited and may become overloaded under bursty local workloads, which calls for cross-node collaboration. In distributed collaborative scenarios without a trusted center, blockchain can provide traceable ledger records for task execution results and related transaction information, but online collaborative task allocation remains challenging under dynamically varying node service capability. Existing methods mainly optimize latency or system utility according to current states and often become less effective when node service capability changes over time. This paper explores the online collaborative task allocation problem in blockchain-assisted MEC and proposes a prediction-driven two-stage collaborative task allocation scheme, PRBM. In the first stage, a multivariate temporal convolutional network predicts the local resource demand of each node in the next slot and accordingly forms the shareable resource amount and the offloading task set. In the second stage, distributed bidirectional matching is used for task allocation, where the Task Publisher selects candidate nodes according to reputation and transmission delay, while the Task Candidate decides whether to accept a task according to unit revenue and resource occupancy. Rewards are then settled according to completion time, and on-chain reputation information is updated accordingly. Through evaluation experiments, we demonstrate that the proposed scheme surpasses several benchmarks in terms of task completion ratio and average makespan.

Keywords: Generative AI, Machine Vision, System Architecture
Abstract: We present DroneDiffusion, a perception-aware diffusion planner for local 3-D UAV trajectory generation that explicitly models and avoids failure zones caused by lens flare, overexposure, and reflective surfaces. A vision-language perceptual safety estimator produces a pixel-aligned reliability map that highlights regions of degraded sensing. A denoising diffusion probabilistic model (DDPM) is conditioned on this map together with the onboard RGB image, goal point, and initial velocity to generate egocentric 3-D trajectories. To enforce obstacle avoidance at inference time without retraining, we apply training-free CHOMP guidance over a 3-D Euclidean signed distance field. The model is trained entirely in simulation and deployed zero-shot on a real drone. Reliability conditioning raises the mean perceptual reliability along generated trajectories to 0.71, compared with 0.63 for a geometric baseline, while guidance halves the collision rate of the learned prior from 14.7% to 7.8%. Perceptual safety conditioning and geometric guidance are therefore complementary: treating sensor reliability as a first-class signal reduces incursions into unreliable regions while maintaining competitive navigation performance.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Real-world video restoration is challenging because degradations such as low-light corruption and rain artifacts vary across frames and often lead to temporal inconsistency, flicker, and unstable visual details. To address these issues, we propose RSTF-DistillDiff, a diffusion-based video restoration framework built upon a residual diffusion model. The proposed method contains three key components. First, we introduce a Residual Scale Block (RSB), which predicts a timestep-dependent scaling factor to adaptively control the magnitude of residual guidance during the diffusion process. This design improves the model’s flexibility under different degradation levels and noise stages. Second, we design a Cross-Frame Attentive Fusion (CFAF) module to aggregate information from the previously restored frame and the current frame, thereby reducing restoration-induced flicker and improving temporal continuity. Third, we employ a structure-preserving distillation strategy to transfer temporal modeling capability from a high-capacity teacher model to a lightweight student model for efficient deployment. Experiments on low-light enhancement and rain-removal benchmarks show that the proposed method achieves competitive restoration quality and favorable temporal stability, while the student model substantially reduces the parameter cost.

Keywords: Human-Machine Interaction, Machine Learning, AI and Applications
Abstract: Automatic speech recognition (ASR) systems struggle with domain-specific terminology in professional settings such as conference presentations, medical consultations, and legal proceedings, where errors critically degrade transcript utility. Existing posthoc correction methods either lack external domain knowledge or risk introducing LLM hallucinations that alter correct transcriptions. We propose DOCENT (Domain-Oriented Conservative Enhancement of Transcripts), a domain-adaptive ASR error correction framework based on Retrieval-Augmented Generation (RAG). Our approach constructs a domain knowledge base from readily available documents, retrieves relevant context for each ASR output, and applies conservative LLM-based correction with three-layer safety constraints: restrictive prompting, diff-based output verification, and retrieval quality filtering. Experiments under a controlled multi-condition protocol (additive white Gaussian noise at five SNR levels from clean to 0 dB) demonstrate: (1) consistent +7–13% Correction Improvement Rate (CIR) across all conditions with an improved-to-harmed ratio up to 13.8:1; (2) correction quality scales with LLM capability (GPT-5.4: +11.24%; GLM-5: +10.06%; Qwen-32B: +9.80%; Gemini Flash-Lite: +4.52%); (3) the conservative safety mechanism maintains positive net improvements under all conditions with minimal harm.

Keywords: Machine Learning, Representation Learning, Deep Learning
Abstract: Harmful meme detection remains a challenging task due to the implicit cross-modal semantics that emerge from the interplay between image and text, rather than from either modality alone. Existing methods typically leverage multimodal pretrained models and cross-modal interaction mechanisms to detect harmful memes. However, they generally rely on entangled representations that mix harmfulness-related causal features with spurious cues, thereby limiting their generalization across different datasets and scenarios. To address this issue, we propose MemeCFD, a novel harmful meme detection approach guided by causal feature disentangled representation learning, which aims to learn truly harmfulness-relevant features while suppressing spurious signals in multimodal representations. Specifically, we first use CLIP to encode memes into unified multimodal representations. Then, we introduce two modality-specific disentanglement modules to separate these representations into causal components that capture genuinely harmfulness-relevant features, and spurious cues that reflect confounding information. Finally, we employ adversarial learning to further promote a more complete separation between causal and spurious features, and perform final prediction exclusively based on the learned disentangled causal representations. Extensive experiments conducted on three benchmark datasets show that MemeCFD outperforms state-of-the-art methods on harmful meme detection task.

Keywords: System Modeling and Control, Reinforcement Learning, Robotic Systems
Abstract: Overactuated four-wheel independently driven and steered(4WID-4WIS) mobile robots suffer from control redundancy coupled with physical constraints, complicating mode selection and coordinated control. We propose a hierarchical hybrid framework fusing deep reinforcement learning with model predictive control, decoupling continuous wheel control into discrete mode decisions at the upper layer and constrained regulation at the lower layer. Given trajectory, motion, and environmental observations, the upper layer switches among Omnidirectional Motion Mode (OMM), Pure Translation Mode (PTM), and Zero-Radius Rotation Mode (ZRM), whereas the lower layer applies single-track equivalent MPC for OMM tracking and geometric mappings for PTM and ZRM. Experiments across three scenarios demonstrate that the proposed method achieves a 100% success rate across all tasks, with superior performance compared to single-mode MPC, rule-based switching, and end-to-end reinforcement learning baselines, especially in constrained scenarios requiring lateral correction and in-place reorientation. Ablation studies further verify the necessity of each motion mode.

Keywords: Machine Vision, Image Processing and Pattern Recognition, Deep Learning
Abstract: Human mesh recovery (HMR) from pressure-sensing bedsheets provides a potential solution for continuous and automatic in-bed healthcare monitoring. Furthermore, this approach offers inherent advantages such as privacy preservation and occlusion robustness. However, the depth ambiguity of pressure signals limits the accuracy of existing HMR methods based on the captured 2D pressure distribution of bed surface. To mitigate this, we derive the key insight that residual pressure signals between adjacent frames physically reflect motion-induced forces, serving as temporal cues to compensate for the lack of information in ambiguous regions. Based on this insight, we present Residual Pressure informed diffusion model for HMR from temporal in-bed pressure images (RP-HMR). Our method designs a diffusion architecture utilizing feature projection strategy and joint-rotation self-attention to model inter-joint structural dependencies, and adopts tailored ControlNet-based fine-tuning to inject these physical constraints for motion guidance. Extensive experiments on the TIP dataset demonstrate that RP-HMR surpasses the state-of-the-art method, achieving 3.84 mm improvement in overall MPJPE and 5.71 mm accuracy improvement in ambiguous regions with limited contact signals, alongside 8 times faster inference speeds.

Keywords: Ethics of AI and Pervasive Systems, AI and Applications, Deep Learning
Abstract: The proliferation of benchmarks for evaluating foundation models has created powerful incentives for manipulation. We study benchmark hacking, the deliberate exploitation of benchmark data during fine-tuning, in the embodied AI domain. We investigate adversarial fine-tuning scenarios with precise control over hacking strategies and intensities. Hacking strategies include explicit training on benchmark data and implicit approaches that indirectly leverage benchmark information through retrieval or synthesis. Hacking intensities are controlled through training duration, data proportion, and exploitation ratio. We develop a detection framework adapting perplexity-based and output concentration measures with reference-free and reference-based aggregation. Our investigation across multiple benchmarks and model architectures reveals a fundamental vulnerability: a sweet spot where exploitation achieves substantial target performance gains with slight in-domain improvements, while maintaining low detection risk and comparable out-of-domain performance. This sweet spot manifests both temporally in early training epochs and in low-to-moderate exploitation levels, exposing systematic vulnerabilities across detection methods and highlighting critical implications for evaluation practices and benchmark design.

Keywords: Reinforcement Learning, Generative AI, Cognitive Computing
Abstract: Multimodal Large Language Models (MLLMs) excel in general scenarios but struggle with specialized situational reasoning, especially fine-grained perception and multi-step logic. While Chain-of-Thought (CoT) prompting enhances reasoning, it suffers from poor adaptability and prompt reformulation across tasks. Thus we propose DeTA, a decoupled two-stage alignment framework that empowers small-scale MLLMs to achieve performance comparable to models up to 8x larger. Stage I injects domain-specific reasoning patterns via Low-Rank Adaptation (LoRA) while preserving general knowledge. Stage II employs Direct Preference Optimization (DPO) for precise CoT alignment using incorrect answers from Stage I as rejected samples, eliminating the need for a reward model. Furthermore, we present an automated pipeline for generating high-quality CoT data with question-guided visual focusing prompts. Trained on a fusion dataset encompassing general, remote sensing, and situational visual question answering tasks, our method achieves an average accuracy of 73.6%, surpassing the flagship models of the same series. Moreover, by incorporating a simple constraint prompt during inference, the model adaptively switches between thinking and non-thinking modes, reducing latency by 21.9% compared to pure long CoT reasoning. Our approach provides an efficient and scalable path for specialized reasoning in resource-constrained environments.

Keywords: Generative AI, Deep Learning, Application of Artificial Intelligence
Abstract: Co-speech gesture generation aims to synthesize human gestures that are temporally aligned with speech and other modalities. Given multimodal inputs, state-of-the-art(SOTA) approaches have generally cast the generation task as an autoregressive or diffusion problem that maps audio and text features to continuous motion sequences. However, these approaches often struggle to capture ultra-long-range temporal dependencies and incur prohibitive computational overhead during inference, further leading to limited generation diversity and style consistency. To address these issues, we propose MamDiff, a novel end-to-end co-speech gesture generation framework. MamDiff employs a Multi-scale Latent Style Mamba (MLS Mamba) module to efficiently capture long-range spatialtemporal dynamics and extract fine-grained stylistic features, and integrates a Diff-GRU generative architecture to seamlessly combine fast autoregressive decoding with highfidelity probabilistic diffusion, significantly improving overall performance. Experimental evaluations demonstrate that MamDiff attains SOTA performance across both the ZEGGS and BEAT benchmarks, thereby confirming the effectiveness of our proposed method.

Keywords: Cyber Systems and Engineering, Knowledge Acquisition, Decision Support Systems
Abstract: Long Time Contributors (LTC) are the backbone of collective intelligence and the sustainable development of open-source communities. Viewed through the lens of social computing, predicting developer retention requires understanding both social interactions and technical capabilities within these socio-technical ecosystems. In this paper, we propose the Socio-KU framework, which bridges this gap by integrating developers' early social behavioral features with programming language Knowledge Units (KU). Based on official Python certifications, we define 15 KUs and extract the Early Developer Knowledge Profile (KULTC_DEV_EXP) from their first 30 days of community engagement, alongside the Initial Project Technical Snapshot (KULTC_PROJ). We validate our approach on 13 open-source Python projects (comprising 8,188,225 commit records, 773,741 pull requests, and 16,859 unique contributors) using a 100-iteration out-of-sample bootstrap to predict 1-year developer retention. Results show that KU features act as a high-value complementary signal to traditional social metrics; the Socio-KU model achieves an AUC of 0.774, outperforming the pure-behavior baseline (AUC=0.756). Feature attribution reveals that the socio-technical alignment between a developer's expertise and the community's knowledge architecture is a critical driver for long-term retention.

Keywords: Game Theory and Multi-Agent Systems, Manufacturing Automation and Systems, Reinforcement Learning
Abstract: Dynamic manufacturing environments require multi-agent systems to coordinate effectively under frequent operational disturbances such as machine failures, urgent job arrivals, and processing time variations. Existing multi-agent reinforcement learning approaches treat each disturbance episode independently, discarding valuable coordination experience that could accelerate future adaptation. In this paper, we propose a Graph-Structured Experiential Memory (GSEM) framework for multi-agent coordination in dynamic manufacturing. The framework encodes historical coordination episodes as heterogeneous relational graphs that capture task dependencies, machine states, and inter-agent collaboration patterns. When a new disturbance occurs, a graph neural network-based retrieval mechanism identifies structurally similar past episodes, enabling experience-guided policy adaptation rather than learning from scratch. Experiments on dynamic flexible job-shop scheduling benchmarks with three disturbance types show that GSEM reduces makespan by 4.1%–10.0% and adaptation time by 33%–38% compared to the strongest memory-augmented baseline, with the advantage increasing under higher disturbance frequency. Ablation studies and cross-disturbance transfer experiments further validate the necessity of graph-structured encoding and similarity-based retrieval, and demonstrate the cross-disturbance generalizability of learned coordination patterns.

Keywords: Cooperative Work in Design, Companion Technology, Team Performance and Training Systems
Abstract: Collaborative design review involves discussions, reviewer feedback, artifact revisions, and revision traces, yet the rationale behind design changes is often fragmented across these heterogeneous sources. Existing work has focused mainly on rationale capture or extraction, with limited support for diagnosing when a decision or revision remains insufficiently justified. We present RationaleGap, a multimodal graph-based framework for rationale gap diagnosis in collaborative design review, reformulated as a collaborative multi-agent system. The framework aligns transcripts, artifact revisions, comments, and metadata into episode-level context, constructs a Cross-Source Design Rationale Graph, and diagnoses four categories of rationale gaps: unsupported decision, missing constraint grounding, missing evidence support, and unresolved reviewer concern. A role-aware repair module further generates actionable follow-up suggestions. Experiments show that the full graph-based model achieves the best performance, reaching 88.05 accuracy and 86.97 Macro-F1, substantially outperforming text-only and graph-free multimodal settings. These results show that rationale gap diagnosis in collaborative review is fundamentally multimodal and benefits strongly from explicit graph-based reasoning.

Keywords: Systems Safety and Security, Machine Learning
Abstract: Reliable network intrusion detection remains challenging due to complex data distributions, diverse attack patterns, and insufficient characterization of prediction credibility. Existing uncertainty-aware methods often rely on the simplified assumption that higher uncertainty directly implies higher prediction risk. However, our analysis shows that, under the current model architecture and experimental setting, misclassified samples are not concentrated in the highest-uncertainty region, but are mainly distributed in the medium-uncertainty interval, indicating that samples near the decision boundary are more critical for model optimization.To exploit this observation, we construct a representation uncertainty-aware network intrusion detection framework based on a Bayesian autoencoder with Monte Carlo sampling, where representation uncertainty is quantified by the variance of similarity between stochastic sample representations and normal prototypes. Building on the resulting uncertainty distribution, we further introduce an adaptive sample reweighting strategy that preserves the emphasis on the medium-uncertainty region while dynamically adjusting sample weights according to their uncertainty characteristics and prediction states. In this way, the model is encouraged to focus more on informative and error-prone samples near the decision boundary, rather than treating all samples in the critical region with the same importance.Experimental results on the NSL-KDD dataset show that the proposed method improves intrusion detection performance over the baseline, increasing the F1-score and Recall. These results indicate that representation uncertainty can provide effective guidance for adaptive sample reweighting and training optimization, thereby helping refine the decision boundary and improve network intrusion detection performance.

Keywords: System Modeling and Control, Autonomous Vehicle, Modeling of Autonomous Systems
Abstract: Accurate aerial interception of intruding drones remains a significant challenge for platforms restricted to low-cost monocular vision. This paper presents an agile interception framework leveraging Image-Based Visual Servoing that incorporates both delay compensation and target-visibility constraints. A pixel-domain Time-Delay Disturbance-Observer Kalman Filter is proposed, utilizing state augmentation and innovation covariance matching to adaptively estimate lumped disturbances and sensing noise. By buffering delayed measurements with forward replay and inter-frame predictive extrapolation, the framework enables delay-free, high-rate prediction despite unknown maneuvers and constant latency. An IMU-assisted Line-of-Sight remapping is utilized to construct attitude-invariant visual errors within a virtual camera frame, ensuring IBVS stability during rapid maneuvers. Furthermore, a visibility-preserving command shaping strategy with barrier-inspired Field-of-View protection is developed to implement hybrid lateral-vertical-yaw corrections, effectively reducing tracking loss during the interception of highly maneuvering targets. The effectiveness and superiority of the proposed methods are validated through pursuit–evasion experiments within a PX4/Gazebo simulation environment, followed by real-world flight tests.

Keywords: Robotic Systems, Deep Learning, Machine Vision
Abstract: Diffusion-based manipulation policies typically rely on teleoperated demonstrations collected separately on each robot, binding the policy to embodiment-specific hardware. We present ZeroTeach, a video-to-action framework that converts human-operated RGB-D demonstrations into embodiment-agnostic kinematic tuples (object point cloud, relative SE(3) end-effector pose, binary gripper) and deploys zero-shot on target robots without target-robot teleoperation, replay, or finetuning. On top of behavior cloning, the diffusion policy adds a per-axis ReLU soft constraint on commanded acceleration as a feasibility prior, mitigating command-side jumps and cascaded drift. On a four-stage dual-block cascaded pick-and-place task, the main real route trains from human RGB-D videos and reaches 54% end-to-end success on the physical RealMan RM65B. An additional rendered RGB-D route reaches 68% / 67% on simulated ARX X5A and RM65B; DP3 and EquiBot baselines reach at most 22% on the physical RealMan.

Keywords:
Abstract: The fascinating collective behaviors of natural swarm systems have inspired extensive studies on configuration generation of drone swarm. In this paper, we propose a synergistic swarm algorithm (SSA) to realize stable spacing configuration and consistent flight of drones. In order to cope with complex mission requirements and achieve safe and fast flight in dense environments, we further propose a flexible obstacle avoidance (FOA) strategy via passive and active environmental adaption. Passive obstacle avoidance algorithm provides drones with self-adaptive forces along drone-obstacle linkages for getting rid of dangerous position and keeping swarm safe. Active obstacle avoidance algorithm provides drones with lateral forces at a certain distance for correcting course of traversal and keeping swarm rapid. We carried out a series of simulation experiments, including swarms of up to 200 drones in mass point model, of up to 50 drones in six degree of freedom model and 4 drones in real experiment. Simulation results illustrated that our strategy and algorithms can ensure fast flight speed and safety of the swarm in dense environments.

Keywords: Deep Learning, Machine Learning, Networking and Decision-Making
Abstract: Graph Neural Networks (GNNs) have achieved remarkable performance in graph representation learning, yet their inherent vulnerability to adversarial attacks poses severe security risks. Especially, black-box node injection attacks have become a major threat to GNNs since they inject malicious nodes without altering the original graph topology. However, they typically decouple the generation of malicious node features and edge connections, thereby resulting in suboptimal attack efficacy under stringent budgets. To address this critical issue, this study proposes a novel Target-aware Interaction-guided Reinforcement learning for Black-box node injection Attacks on GNNs (TIRBA), which formulates the attack as a Markov Decision Process and jointly optimizes node feature generation and edge construction in a heterogeneous action space. Firstly, TIRBA designs a target-aware interaction encoder to fuse information of node features and edges. Further, it introduces a class-center guidance mechanism to utilize prior class distribution information, thereby guiding efficient exploration of the high-dimensional feature space. Finally, a topology difference-aware state value evaluation is adopted to explicitly capture local structural anomalies caused by injected nodes, thereby stabilizing the reinforcement learning training process. Experimental results demonstrate that the proposed TIRBA significantly outperforms state-of-the-art black-box node injection attack methods.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Machine Vision
Abstract: To address the dual challenges of microscopic feature loss and feature collapse caused by long-tailed distributions in the ethnic recognition of minority clothing images, we propose TSRC-ViT(Texture-Semantic Resonance and Calibration ViT), a novel recognition model based on adaptive resonance and generative feature calibration. The model utilizes the DT-SAR module to dynamically fuse shallow microscopic textures, such as embroidery, with deep macroscopic semantics to enhance feature discriminability. Furthermore, it introduces the CG-GFC framework, which explicitly generates hallucinated samples using the statistical priors of many-shot categories and reshapes the decision boundary alongside logit adjustment to fundamentally overcome the prediction inertia for rare ethnic categories. Experiments on a custom large dataset covering 55 ethnic minorities demonstrate that TSRC-ViT improves few-shot accuracy, Group-Average accuracy, and macro recall while introducing minimal computational overhead. The model achieves an optimal balance among parameters, computation, and accuracy, providing robust technical support for the preventive protection of digital humanities resources related to ethnic minority clothing.

Keywords: Intelligent Transportation Systems, Transfer Learning, Representation Learning
Abstract: While domain adaptation (DA) methods like adversarial training (e.g., DANN) and moment matching (e.g., MMD) have been widely adopted to address data scarcity in spatio-temporal forecasting, they fail to separate causal invariant patterns of urban systems from spurious correlations caused by transient disturbances. By conflating these two sources of variation, such methods erroneously transfer noise-driven statistical similarities as though they were stable causal knowledge, ultimately leading to negative transfer. We argue that robust cross-domain prediction requires identifying invariant spatio-temporal dependencies, rather than aligning statistically similar but potentially spurious patterns. These invariant spatio-temporal dependencies are mechanisms whose predictive relationship with the target remains stable across environments cite{arjovsky2019invariant}. Based on this insight, we propose the Invariant Causal Representation Spatio-Temporal Graph Neural Network (ICR-STGNN), a principled framework that disentangles structural invariance from contextual noise through dual-stream representation learning. By decomposing latent dynamics into commonality (capturing persistent urban regularities such as road topology and habitual flow) and specificity (modeling transient disruptions like weather anomalies or incidents), ICR-STGNN enforces invariance via a multi-domain consistency objective derived from causal invariance principles. The resulting invariant representations generalize reliably to unseen domains with minimal supervision. Extensive experiments on real-world traffic datasets demonstrate that our method achieves state-of-the-art performance under diverse distribution shift scenarios.

Keywords: Generative AI, AI and Applications, Smart Buildings, Smart Cities and Infrastructures
Abstract: The rapid advancement of artificial intelligence (AI) technologies, particularly large language models (LLMs), has opened new avenues for automating complex tasks across various industries. In the field of power infrastructure con struction, generating construction plans remains a highly manual, time-consuming process, often prone to inefficien cies and inconsistencies. This paper proposes an LLM-based framework for the automatic generation of construction plans, incorporating domain-specific knowledge retrieval, multi-task constrained generation, and multi-dimensional reward model ing. The framework utilizes a Transformer-based dual-encoder architecture to perform entity extraction and scenario intent classification, followed by a retrieval-augmented generation (RAG) module that integrates relevant engineering standards and technical documentation. The multi-task generation module synthesizes key construction elements—such as procedures, equipment lists, safety measures, and risk assessments—while strictly enforcing domain-specific constraints to ensure consis tency and compliance. A multi-dimensional reward model is introduced to evaluate the generated plans across several di mensions, including terminology accuracy, process compliance, safety completeness, and operability. Extensive experiments demonstrate that the proposed system significantly improves generation efficiency, text quality, and domain compliance, outperforming baseline methods in both quantitative metrics and expert evaluations. This work highlights the potential of integrating LLMs with domain-adapted retrieval and multi task generation to advance automation and intelligence in the power infrastructure sector.

Keywords: Deep Learning, Image Processing and Pattern Recognition, Neural Networks and their Applications
Abstract: Accurate segmentation of the anterior visual pathway (AVP) from magnetic resonance imaging (MRI) is critical for surgical planning and various medical interventions. Conventional convolutions effectively capture local features but struggle with modeling long-range dependencies. To address this, we propose a large-kernel driven multi-scale context alternating fusion model for fully automatic AVP segmentation. By incorporating a modified large-kernel depth-wise separable convolution block, the framework expands the receptive field and enhances long-range dependency modeling. An interleaved context extraction encoder further enables the fusion of multi-scale and multi-level long-range dependencies with local details. We constructed the AVP-119 dataset by collecting, filtering, and annotating 119 brain MRI scans. Extensive experiments and ablation studies demonstrate that our method outperforms state-of-the-art segmentation approaches on AVP-119 (Dice: 90.10%)

Keywords: AIoT, Application of Artificial Intelligence, Representation Learning
Abstract: Human activity recognition serves as the foundation for various emerging applications. In recent years, researchers have used collaborative sensing of multi-source sensors to capture complex and dynamic human activities. However, multimodal human activity sensing typically encounters highly heterogeneous data across modalities and label scarcity, resulting in an application gap between existing solutions and real-world needs. In this paper, we propose CLMM, a general contrastive learning framework for human activity recognition that achieves effective multimodal recognition with limited labeled data. CLMM employs a novel two-stage training strategy. In the first stage, CLMM employs a CNN-DiffTransformer encoder to capture cross-modal shared information by extracting local and global features. Meanwhile, a hard-positive samples weighting algorithm enhances gradient propagation to reinforce shared learning. In the second stage, a dual-branch architecture combining quality-guided attention and bidirectional gated units captures modality-specific information, while a primary–auxiliary collaborative training strategy fuses both shared and modality-specific information. Experimental results on three public datasets demonstrate that CLMM significantly improves state-of-the-art baselines in both recognition accuracy and convergence performance.

Keywords: Human-Machine Interface
Abstract: 文本驱动生成的可编辑参数CAD 模型仍面临三个持续的挑战：（i）空间 定位很难——起点和草图平面 法线/取向通常隐含于语言中， 产生不可执行序列;（ii） 弱耦合 数值参数与几何/约束之间的领先 在长期施工过程中误差累积;以及 （三）广泛的教学 细致度——从抽象描述到专家级 规范——缺乏一致的监督和 评估。动力来源 我们分为两条工作线，介绍SLM-CAD。在数据方面， 灵感来自Text2CAD和CAD-GPT的注释 食谱—— 主要采用视觉语言模型（VLM）来推导 DeepCAD 的抽象级参数描述——我们 装置 一个四级注释流水线：多视图VLM形状 最小元数据重建→描述 →细粒度自然语言教学（NLI） 四级指令→生成 准备。该流水线提供从抽象到的监督 专家 在减少幻觉和改善的同时，水平下降 可执行性。在&

Keywords: Image Processing and Pattern Recognition, AI and Applications, Machine Learning
Abstract: Locally advanced rectal cancer (LARC) represents a significant cause of cancer-related mortality globally. The accurate preoperative prediction of therapeutic response, particularly the achievement of pathologic complete re-sponse (pCR), is critical for optimizing treatment strategies and improving patient outcomes. While histopathological examination serves as the gold standard for LARC diagnosis and prognosis, the analysis of whole-slide im-ages (WSIs) presents substantial challenges for multiple instance learning (MIL) due to their gigapixel resolution. Current MIL approaches typically represent WSIs as instance-bag configurations, yet they face limitations in effectively capturing inter-instance interactions, often relying on approxima-tions like linear attention mechanisms. To overcome these limitations, we propose MamDG, a novel analytical framework that integrates a selective structured state-space model with dynamic directed graph representations. This approach effectively models global instance dependencies while mini-mizing topological sequence traversal, thereby preserving essential interin-stance relationships. In this retrospective observational study, we analyzed data from 118 LARC patients who underwent neoadjuvant chemotherapy at a tertiary care institution. Our experimental results demonstrate that MamDG significantly outperforms existing frameworks in preoperative therapeutic re-sponse prediction tasks.

Keywords: Deep Learning, Image Processing and Pattern Recognition, Medical Informatics
Abstract: Medical image analysis often involves different lesions whose diagnosis relies on diverse visual cues, such as subtle texture changes, irregular boundaries, and complex spatial arrangements. For example, in lymph node assessment, central necrosis primarily depends on internal texture, while nodal matting relies on structural fusion. Existing medical image classification methods commonly adopt hybrid architectures combining convolutional neural networks and Transformers to capture complementary visual patterns. However, these architectures typically aggregate heterogeneous visual cues into a single feature representation, resulting in implicit entanglement and limiting their ability to explicitly model diverse lesion characteristics and multi-view visual patterns. To address these challenges, we propose FAMMA-Net, a Frequency-Aware Multi-view Mixture-of-Adapters Network for multi-visual-cue lesion recognition. A Deep Frequency Filtering (DFF) module generates a sample-adaptive frequency view by dynamically selecting informative frequency components for each input, enabling the model to retain lesion-relevant high-frequency cues. A Frequency-Guided Enhancement (FGE) module adaptively refines embeddings to emphasize lesion-sensitive patterns. A Multi-view Mixture-of-Adapters (Mv-MoA) integrates the spatial, frequency, and enhanced views, allowing different adapters to specialize in capturing distinct visual cues through feature modulation. Extensive experiments on multiple medical imaging datasets demonstrate that FAMMA-Net achieves robust and reliable performance for multi-visual-cue lesion recognition.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Machine Learning
Abstract: FDCP-Mamba combines frequency-driven detail aggregation and center-guided prototype calibration for high-resolution remote sensing semantic segmentation, achieving competitive results on LoveDA and ISPRS Vaihingen.

Keywords: Affective Computing, Image Processing and Pattern Recognition, AI and Applications
Abstract: 识别（MER）旨在揭示微妙的 1644;短暂的面部表情 2290;现有方法通常߶ ;1;赖 关于齐次特征提 1462;和忽略 非对称起始——偏移& #21160;力学固有于 微表达式，导致 2312;上的不稳定表示 数据集有限。为 5299;决这些限制，ৎ ;5;们提出建议 PMNet，一个相位感௽ ;3;协同的三流网络& #65306;a 多尺度分层空间 1464;换器（HST） 静态表征学习，9968;种相位感知协ࡧ ;6;网络（PAS-Net），明 830;建模起始点--偏ĸ ;27;与交叉注意力的 1160;态，以及通道ࣜ ;6;强 卷积神经网络（CEC -Net）能够精炼 低层次的视觉提 1034;和频道专注。ߎ ;1;像 将时间线索压缩 0026;一的方法 表征、相位感知 1327;同作用保持起ࣺ ;7;和偏移 作为独立的阶段 1644;模型，它们互ඣ ;7;通过门控跨阶段 7880;意力进行互动Ӎ ;0;在 遗漏一个主体（LOS O）交叉验证策略&# 65292;对公共基准

Keywords: Expert and Knowledge-Based Systems, Application of Artificial Intelligence, Generative AI
Abstract: Adaptive retrieval-augmented generation (RAG) requires estimating when a language model needs external knowledge, yet existing uncertainty methods---within-query semantic entropy and attention-based signals---condition on a fixed retrieval context and miss how answer stability varies across different retrieved documents. We propose Cross-Query Semantic Entropy (CQ-SE), a training-free method that generates semantically equivalent query paraphrases, retrieves documents independently for each, and measures answer disagreement across these diverse retrieval contexts. This cross-query signal captures retrieval-dependent uncertainty that within-query sampling cannot detect. Evaluated on five QA benchmarks at two model scales (Qwen2.5-7B/72B), CQ-SE achieves 0.636--0.723 average AUROC for retrieval necessity detection, outperforming SUGAR and INTRYGUE by 18--37 percentage points, with uniquely positive scaling behavior. Ablation analysis confirms that each pipeline component contributes incrementally, with cross-query perturbation providing the largest gain, and mechanism analysis confirms that retrieval variability significantly correlates with CQ-SE's advantage.

Keywords: Image Processing and Pattern Recognition, Human-Machine Interaction, Deep Learning
Abstract: 随着体育分析深度学习的发展，快速竞技体育的早期预测变得具有挑战性。现有的羽毛球动作识别与预测研究主要关注已完成的动作，且常依赖单一信息源或静态坐标，这会失去动态特征。为解决这些限制，采用了一种积分沫球轨迹的双分支预测方法

并开发骨骼姿态序列。该方法依赖摄像机角度、位置和数字，并通过手动标注羽毛球拉力赛视频构建数据集，提取从羽毛球越网到球拍准备的关键时间段。YOLOv8和OpenPose提取羽毛球轨迹和骨架关键点。基于这些数据，计算衍生特征以提升预测准确性，缩放因子则减少摄像机角度和球员体型的影响。应用了帧差分和主成分分析，卷积神经网络（CNN）和双向长短期记忆网络（Bi-LSTM）则学习局部动态和长

Keywords: Control of Uncertain Systems, System Modeling and Control, Complex Network
Abstract: Within this article, the asynchronous event-triggered exponential stabilization of stochastic large-scale networks driven by G-Brownian motion (G-SLSNs) within G-expectation framework (GEF) is investigated, where an asynchronous Denial-of-Service attacks (ADoSA) is considered. Diverging from the existing event-triggered stochastic stabilization results, uncertainties in probabilities and models are factored to more accurately describe real-world operational scenarios. Recognizing that the specific positive homogeneity and sub-additivity of GEF render current stability theories ineffective, we establish a piecewise DoS-dependent G-Lyapunov functional (G-PDLF) parameterized by G-noise to jointly capture asynchronous DoS switching and G-scenario stochastic dynamics, alongside a novel execution error estimation method adapted to the mathematical properties of GEF, which yields less stringent stability conditions. By the Lyapunov functional framework and analysis method based on G-stochastic characteristics, exponential stabilization criteria for event-triggered G-SLSNs under ADoSA are derived. Lastly, the theoretical outcomes are demonstrated numerically via a single-link flexible joint manipulator system.

Keywords: Technology Assessment, Information Systems for Design, Networking and Decision-Making
Abstract: 本文研究了高速链路和内存受限条件下，软件定义网络（SDN）中的轻量级 DDoS 检测问题。不同于维护完整的逐流状态，本文提出了 CxSketch，这是一种候选流保留结构，通过结合主表和辅助表，在严重哈希冲突下减少潜在大流被过早驱逐的情况。CxSketch 利用流量大小和到达连续性来指导替换与提升决策，从而为在线检测提供更稳定的输入。本文进一步将 CxSketch 与轻量级级联检测器结合，在对齐的时间窗口下区分洪泛攻击和慢速率攻击。CxSketch 面向软件定义的转发与监测架构，例如 OVS、DPDK、eBPF/XDP 路由器以及控制器侧遥测采集器，而不是面向线速可编程 ASIC 流水线。实验结果表明，在小内存预算下，CxSketch 在 Top-K 恢复和重流检测方面表现良好。在 4KB 设置下，它取得了 0.94 的 Top-K

Keywords: Generative AI, Expert and Knowledge-Based Systems, Game Theory and Multi-Agent Systems
Abstract: Retrieval-Augmented Generation (RAG) mitigates hallucinations in Large Language Models (LLMs) by grounding responses in external knowledge, yet erroneous or biased retrieval can still mislead generation, compounding hallucinations---a phenomenon termed Hallucination on Hallucination. The recently proposed Debate-Augmented RAG (DRAG) framework addresses this issue by integrating Multi-Agent Debate (MAD) into both retrieval and generation. However, DRAG still faces three challenges: single-answer generation introduces high randomness, response debate can drift away from the original question, and no systematic contradiction detection mechanism is provided. We propose DRAG-Enhanced, a contradiction-aware multi-candidate verification framework that extends DRAG through a six-stage pipeline centered on candidate diversification and downstream verification. The framework combines: (1) 2-by-2 factorial candidate generation via prompt style and sampling temperature; (2) token-level ensemble for fine-grained aggregation across candidates; (3) judge-based answer selection supported by contradiction-aware verification; and (4) a conservative anti-loop feedback signal for a small subset of high-conflict cases. Experiments on three benchmark datasets show that DRAG-Enhanced consistently improves over DRAG in Exact Match (EM), achieving the best observed EM results on Natural Questions and HotpotQA while remaining competitive on PopQA. In particular, it improves EM by 13.34 points on Natural Questions and by 3.33 points on HotpotQA over DRAG.

Keywords: Affective Computing, Quantum Machine Learning, Deep Learning
Abstract: Modeling with quantum theory shows promise in NLP, yet most rely on fixed interference terms and lack dynamic adaptive mechanisms, limiting their ability to handle varying sentence-level semantic ambiguity. To this end, we propose QDA, an end-to-end Quantum-inspired Dynamic Adaptive framework for textual sentiment analysis. It is the first to incorporate dynamic adaptive interference, which enables the effective resolution of semantic ambiguity via three dedicated modules. First, Quantum-inspired Embedding integrates pre-trained models with quantum encoding to capture rich semantic-syntactic information. Second, Quantum-inspired Dynamic Adaptive Interference adjusts interference strength according to sentence-level ambiguity, enabling flexible quantum feature fusion. Third, Quantum-inspired Sentiment Measurement performs measurements on quantum fused features to predict accurate and interpretable results. Finally, experiments on five benchmarks show that our method achieves SOTA performance with 94.00% average accuracy, outperforming existing methods. Additional analyses also support the demonstrated effectiveness and interpretability of the proposed method.