Keywords: BMI Emerging Applications, Active BMIs, Brain-based Information Communications
Abstract: This study investigates the feasibility of using steady-state motion visual evoked potentials (SSMVEPs) as subject-specific biometric signatures under a single-stimulus paradigm. In contrast to conventional SSVEP-based brain computer interface approaches that rely on multi-target stimulus decoding, the proposed framework presents a single motion based visual stimulus and examines whether the resulting neural responses contain sufficient individual-specific information for identification. The stimulus consists of a three-layer concentric motion pattern with controllable layer-specific parameters. Two configurations were evaluated, differing in their motion characteristics. EEG signals were recorded from occipital and parieto-occipital regions under stimulation frequencies of 4, 6, 10, and 15 Hz. Eight participants completed two recording sessions separated by approximately 20 minutes. In each session, data were collected for both stimulus configurations and all stimulation frequencies. A frequency-specific identification framework was adopted, where separate models were trained for each frequency. Results show that SSMVEP responses contain subject-specific information, with performance varying across frequencies and configurations. These findings highlight the importance of frequency selection and suggest that motion-based stimulation provides a comfortable alternative to flicker-based paradigms.

Keywords: Active BMIs, Brain-Computer Interfaces, Brain-based Information Communications
Abstract: Steady-State Visually Evoked Potential (SSVEP)-based Brain–Computer Interfaces (BCIs) rely on frequency tagged visual targets for control. This pilot study investigates whether augmenting conventional SSVEP targets with a dynamic visual overlay derived from Conway’s Game of Life affects online BCI performance and subjective user experience. Participants completed standardized spelling tasks under two conditions: (1) conventional flickering stimuli and (2) identical flickering stimuli augmented with a Game-of-Life overlay. Stimulation frequencies and visual parameters were held constant across conditions. EEG recorded from occipital channels was classified using Filter Bank Canonical Correlation Analysis (FBCCA). Performance was evaluated using classification accuracy and Information Transfer Rate (ITR), alongside subjective attentional and fatigue ratings. Across all tasks, accuracy was 97.4% (95% CI: [94.4, 100]) without overlay and 95.6% (95% CI: [92.4, 98.8]) with overlay; ITR was 40.95 (95% CI: [29.19, 52.70]) bits/min without overlay and 34.53 (95% CI: [21.33, 47.73]) bits/min with overlay. Subjective preference was split evenly between conditions (50% vs. 50%), tiredness increased from 2.33 ± 0.82 to 4.00 ± 1.26 on a 6-point Likert scale.

Keywords: BMI Emerging Applications, Active BMIs, Brain-based Information Communications
Abstract: This paper presents an SSVEP-based brain-computer interface for solving chess puzzles on a custom illuminated physical chessboard. The system combines an 8x8 LED stimulus matrix, transparent chess pieces, and a Python-controlled interaction loop that assigns stimulation frequencies only to chess-legal squares. EEG signals recorded from occipital and parietal electrodes are decoded using filter bank canonical correlation analysis (FBCCA). The current board state determines which squares are stimulated and restricts the set of candidate outputs considered during classification. In an online pilot study, six participants solved mate-in-two puzzles selected from the Lichess database. The system achieved a participant-level mean operational accuracy of 92.7% with a 95% confidence interval of [85.8, 99.6]%. The mean practical puzzle-completion time was 102.52 s, with a 95% confidence interval of [45.39, 159.65] s. Questionnaire results indicated manageable flickering discomfort, a moderate increase in fatigue, and generally positive usability ratings. These results shows the feasibility of tangible SSVEP-based chess interaction and suggest that restricting stimulation to chess-legal moves can support reliable BCI control in a physical game environment.

Keywords: Neural Networks and their Applications, Expert and Knowledge-Based Systems, Multimedia Systems
Abstract: Parameter-efficient fine-tuning (PEFT) has garnered substantial attention owing to the expansion of pre-trained model sizes and the necessity to fine-tune them for enhanced downstream performance. Currently, adapter method has become one of the mainstream PEFT choices, due to its plug-and-play nature and strong cross-task adaptability. However, when the number of trainable parameters is extremely small and downstream samples are scarce, the optimization space is highly constrained and the loss landscape tends to fall into sharp local minima, which causes a sharp drop in generalization performance. Here, we identify insufficient diversity as a key factor behind performance degradation and propose the Dual-Adapter framework to address this issue. By introducing perturbation-based diversification, our method enhances optimization landscape exploration and mitigates the risk of convergence to sharp minima. Specifically, Dual-Adapter employs two structurally identical yet independently initialized adapter branches, which simultaneously pursue distinct optimization paths during training. This dual-path design enables consistent generation of diverse gradients, promoting more robust and stable convergence.Experiments on 11 datasets achieve an average improvement of +9.9%, outperforming both classical and state-of-the-art fine-tuning methods.

Keywords: Image Processing and Pattern Recognition, Application of Artificial Intelligence, Deep Learning
Abstract: Supervised denoising models suffer from domain shift, while zero-shot alternatives lack high fidelity. We propose Probabilistic Dual-Perception Denoising (PDPD), a zero-shot framework utilizing Gaussian denoisers as internal priors to bridge this gap. Our Siamese Dual-Path Topology synergizes an "occluded" path for noise rejection with a "visible" path for detail recovery. To handle complex non-Gaussian noise without external data, Statistical Manifold Rectification (SMR) enforces Gaussian latent statistics, and Bayesian Visibility Distillation (BVD) distinguishes textures from heteroscedastic noise. Extensive experiments across synthetic shifts and five real-world datasets (e.g., SIDD+, Nam, DND) demonstrate that PDPD achieves state-of-the-art zero-shot performance (37.11 dB on SIDD+), closely rivaling supervised baselines.

Keywords: Decision Support Systems, Adaptive Systems, Conflict Resolution
Abstract: Conventionally, decision authority in human-robot collaboration is dominated either by human reasoning or by autonomous robot (AI agent) control with limited human oversight. With the increasing need for adaptive and reliable decision-making in complex robotic environments, there is growing interest in joint inference, where human and AI judgments are combined. However, current joint inference methods typically rely on confidence-only heuristics such as Maximum Confidence Slating (MCS), which overlook task context and the reliability of confidence itself. In this work, we present an adaptive framework for human-AI joint inference that goes beyond confidence-only selection by incorporating task features and agents’ metacognitive signals. Our method uses Contextual Bandits to dynamically combine human and AI judgments based on confidence, task difficulty, and metacognitive traits, including sensitivity (AUROC2 measures) and calibration (ECE measures). We evaluate our approach on classifying differential operational delays in multi-robot teleoperation, where subtle delays timing differences can only be inferred through temporal interaction. Experimental results show consistent accuracy gains over confidence-only baselines, with metacognitive features providing further improvements, particularly in difficult tasks. We also introduce a Latent Variable Bandit variant that remains effective when task difficulty is unobservable. This work provides the first principled framework for adaptive human-AI joint inference in spatiotemporal robotic tasks, highlighting how metacognitive signals can improve collaboration between humans and embodied AI systems.

Keywords: Virtual, Augmented, Mixed Reality, Human-Computer Interaction, Machine Learning
Abstract: A-mode ultrasound (US) is a promising sensing modality for Virtual Reality (VR) interaction, as it enables the mapping of muscular activity into control commands while retaining the benefits of wearable sensing. However, existing approaches still face limitations in terms of wearability and interaction complexity, often relying on external hardware such as cameras. In this work, we propose a fully wearable multimodal interface for real-time VR-interaction, based on concurrent US and inertial (accelerometry) sensing from the forearm and upper arm. The system is built on the WULPUS platform and integrates an end-to-end software framework for real-time acquisition, visualization, and communication with a Unity-based VR environment. A multimodal learning pipeline is introduced for concurrent hand pose and forearm position estimation in 2D space. The interface is evaluated through offline and online experiments with five subjects, during the execution of three functional tasks: cylinder grasping (gross motor) and relocation, marble pinching (fine motor) and relocation, and liquid pouring. For offline experiments, we collect 5 acquisition sessions across multiple days, achieving an average inter-session accuracy across subjects of 80±6% for hand pose estimation and 77±7% for forearm position estimation. Online validation with minimal fine-tuning (5 min) demonstrates success rates of 92.0±16.0%, 88.0±9.8%, and 96.0±8.0% for the three tasks, respectively. With a power consumption of only 19.9 mW, our system enables more than 2.5 days of continuous use on a small 350 mAh LiPo battery without the need for recharge, enabling truly wearable, multimodal, and functionally meaningful VR interaction.

Keywords: Affective Computing, Human-centered Learning, Machine Learning
Abstract: Affective computing tasks, such as stress and drowsiness detection, are challenging due to complex multimodal dependencies and inherent inter-subject variability. While addressing these challenges typically requires high-capacity, personalized ensembles, deploying such complex models for everyday edge use is impractical due to their high computational costs. Conversely, simply compressing these models often degrades accuracy, as smaller networks struggle to adapt to diverse users and frequently latch onto user-specific noise. To address this, we propose a framework that first distills knowledge from a heavy, multi-teacher ensemble into a compact student network. To handle inter-subject variability, this student is equipped with an identity-aware gating mechanism that dynamically adjusts modality importance based on the user. Finally, we introduce an active optimization loop using local explanations to identify and penalize user-specific confounding features during fine-tuning, while reinforcing reliable signals. Evaluated on the UL-DD and EmpathicSchool datasets, the optimized student model achieves up to a 4.6× inference speedup over the teacher ensemble while improving accuracy to 88.00% and 87.48%, respectively. Notably, this lightweight model outperforms both the high-capacity teacher ensemble and heavier attention-based baselines. Ablation studies confirm these gains stem directly from the multi-teacher distillation and targeted feature suppression. These findings demonstrate that explainability can move beyond passive analysis to serve as active regularization, improving the efficiency and inter-subject consistency of affective computing models.

Keywords: Brain-Computer Interfaces
Abstract: Motor imagery (MI) brain-computer interface (BCI) systems often suffer from reduced reliability due to involuntary command triggering during intentional non-control (INC) periods. This study proposes an unsupervised gating framework designed to prevent unintended command triggering. The system evaluates the spatial focalisation of electroencephalography (EEG) signals modelled through a gaussian mixture models (GMM), filtering out neural samples that lack the distinct neurophysiological signature of motor intent before they reach the command integrator. Real-time closed-loop results demonstrate that the introduction of Spatial Focalisation Gating (SFG) significantly increases INC Accuracy from 24% to 76% compared to traditional architecture, while maintaining a high Trial Accuracy of 92.7%. Despite a moderate increase in command latency, the SFG framework ensures superior synchronous safety, providing a robust and predictable interface for real-world applications.

Keywords: Brain-Computer Interfaces, Active BMIs
Abstract: Speech imagery is attractive as a brain-computer interface paradigm for communication because it is endogenous and intrinsically linguistic. Yet despite growing interest, its dominant scalp-EEG spatiotemporal characteristics remain poorly characterized. Here, we asked how speech imagery appears in scalp EEG and compared it against finger motor imagery. Using a within-subject dataset containing speech imagery, finger motor imagery, and no-task trials recorded under the same trial structure, we analyzed band-power dynamics across channels and time. Finger motor imagery showed the expected contralateral mu/alpha and low-beta desynchronization over sensorimotor areas, whereas speech imagery showed a weaker, more distributed alpha-dominant increase. After normalization to each condition's own post-trial interval, the speech-related alpha increase changed only modestly after cue onset, indicating that much of the speech-versus-no-task difference was already present during the instruction period. A classifier discriminating imagery from no-task reached mean balanced accuracies of 0.563 pm 0.072 for speech imagery and 0.718 pm 0.127 for motor imagery, with a stronger alpha/beta dependence for motor imagery than for speech imagery. Together, these results provide a clearer group-level characterization of speech imagery in scalp EEG and indicate that its dominant spatiotemporal pattern differs from that of finger motor imagery and is more consistent with substantial non-articulatory task-related contributions than with a clear articulatory-motor analogue.

Keywords: Brain-Computer Interfaces, Computational Intelligence, Machine Learning
Abstract: Steady-state visual evoked potential (SSVEP) based brain–computer interfaces (BCIs) enable high-speed communication by decoding frequency-specific responses from electroencephalography (EEG) signals. Reliable frequency recognition becomes challenging for short gaze durations due to low signal-to-noise ratio (SNR). In this study, Singular Value Decomposition (SVD) is employed as a dimensionality reduction technique to enhance correlation-based SSVEP classification. The proposed approach applies SVD-based preprocessing to extract dominant signal components from multichannel EEG recordings before classification. The method is evaluated on a benchmark SSVEP dataset containing recordings from 35 subjects and 40 stimulation targets. Canonical Correlation Analysis (CCA) and Filter Bank CCA (FBCCA) are compared with their SVD-enhanced variants across multiple gaze durations. The results show that the SVD-enhanced FBCCA method achieves the best performance, reaching a maximum classification accuracy of 95.38 % and an information transfer rate (ITR) of 112.57 bits/min at an optimal gaze duration of 1.14 s. These findings indicate that the SVD-based approach can effectively suppress noise and improve frequency recognition performance in SSVEP-based BCIs.

Keywords: Brain-Computer Interfaces, Deep Learning, Neural Networks and their Applications
Abstract: EEG is a non-invasive method for recording electrical activity in the brain and is widely utilized in brain-computer interfaces (BCI) to decode motor imagery (MI) signals. However, due to the complex long-term temporal dependencies inherent in MI signals and the difficulty in capturing cross-frequency interaction features, decoding MI-EEG signals remains a challenging task. To address these challenges, we propose IFANet, a novel, comprehensive, and efficient model. This model integrates interactions between different frequencies with the attention mechanism within the Transformer architecture. This approach models cross-frequency interactions while effectively capturing both local and global dependencies, thereby enhancing feature extraction and decoding accuracy. Experimental results on three publicly accessible datasets (BCIC-IV-2a, BCIC-IV-2b, and the High-Gamma Dataset) demonstrate that IFANet achieves superior classification accuracy compared to numerous state-of-the-art models. The efficacy and interpretability of the model are further validated through ablation studies and visualization techniques, highlighting its robust decoding capabilities.

Keywords: Brain-Computer Interfaces, Active BMIs, BMI Emerging Applications
Abstract: Robust classification of motor imagery electroencephalography (MI-EEG) signals is a fundamental challenge in brain-computer interface (BCI) research, largely owing to the non-stationarity and cross-subject heterogeneity of these neural recordings. This paper proposes MTCANet, a deep learning framework that integrates multi-scale convolution with hierarchical attention mechanisms to strengthen MI-EEG decoding. The network architecture is built around three cooperating modules: (i) Multi-Scale Convolution with Multi-Head Self-Attention (MSCA) blocks that extract representations at different scales, (ii) a Temporal Convolutional Module (TCM) that fuses and compresses these multi-scale features while modeling both short-range and long-range dependencies, and (iii) an Efficient Channel and Convolutional Block Attention Module (ECBAM) that dynamically weights and combines high-dimensional features from different scales to integrate signal features and enhance decoding performance. Training is guided by a joint objective that combines cross-entropy loss with center loss, promoting both inter-class separation and intra-class compactness in the learned embedding space. Evaluated on the BCIC-IV-2a dataset, MTCANet attains a classification accuracy of 86.61%, which achieves the best overall performance. Systematic ablation experiments confirm that each architectural component yields a meaningful contribution. These results provide valuable insights for exploring how to effectively fuse multiple types of feature information and how to utilize attention mechanisms more efficiently for practical MI-EEG decoding.

Keywords: Brain-Computer Interfaces, Deep Learning, Representation Learning
Abstract: EEG-based motor imagery classification has been optimized almost exclusively for accuracy, leaving the internal geometry of backbone embedding spaces unmeasured and uncontrolled. We introduce Intra-Class Representation Regularization (ICRR), a loss-level intervention that adds a squared centroid-deviation term to standard cross-entropy training, enforcing intra-class compactness in the backbone embedding space without architectural modification or additional inference cost. Evaluated on the BNCI 2014-001 four-class motor imagery benchmark under nine-fold Leave-One-Subject-Out cross-validation with DeepConvNet as the backbone, ICRR reduces the Embedding Stability Score — mean L2 distance from each embedding to its class centroid, averaged over classes — by 23.2% at the selected λ = 0.25 (d = 5.641, p < 0.0001, N = 9), with no statistically significant accuracy degradation (Δ = −0.81 pp, p = 0.490). These results demonstrate that embedding geometry and classification accuracy are independently optimizable properties of the same model, and that intra-class compactness can be explicitly controlled through a single auxiliary loss term without perturbing classification performance.

Keywords: Robotic Systems, Machine Vision, Environmental Sensing
Abstract: The deployment of unmanned aerial vehicle (UAV) in applications such as agriculture, disaster reconnaissance, and search and rescue necessitates robust and accurate localization. However, traditional localization techniques suffer from significant performance degradation in signal-limited environments. In addition, existing visual approaches often fail to provide reliable and precise pose estimation during the critical initialization stage, particularly for the nadir-view UAVs. To overcome these challenges, this paper constructs a new visual-inertial odometry (VIO) framework. The core of our approach is a dynamic pose estimation strategy driven by motion pattern awareness, which analyzes the UAV motion and scene geometry to select the most suitable estimation technique adaptively. Notably, the proposed strategy enables seamless transitions between planar and non- planar environments without the needs of additional hardware. Experiments conducted on several public datasets demonstrate the effectiveness of our approach. The results have shown that the proposed framework achieves superior localization accuracy and significantly improves overall VIO performance compared to state-of-the-art techniques. Code and datasets are available at https://github.com/jamesswim/VINS-Fusion.

Keywords: Control of Uncertain Systems, Mechatronics, Adaptive Systems
Abstract: This study presents an adaptive neural sliding- mode controller designed for rigid robotic manipulators. The controller comprises a sliding disturbance observer (SDOB) and a neural controller. Utilizing the universal function approximation capability of a neural network, the neural controller adaptively models and compensates for unknown nonlinear dynamics and disturbances. The SDOB enhances tracking performance during the transient learning phase of the neural controller, improves system stability via a sliding-mode technique, and effectively reduces chattering. A key aspect of this learning approach is the utilization of a signal that corresponds to the error of disturbance cancellation. It enables a straightforward training process for the neural network, as opposed to relying on tracking error, which includes reference signals that are irrelevant to the disturbances. Real-time implementation on a physical robotic manipulator has been carried out, and the experimental results clearly demonstrate the proposed scheme’s viability.

Keywords: Fuzzy Systems and their applications, System Modeling and Control, Intelligent Industrial Systems
Abstract: This paper proposes the Recurrent Positive Polynomial Fuzzy Stacked Broad Learning System for Adaptive MIMO-PID Control (RPFSBLS-APID), which is a unified framework for online MIMO system identification and setpoint tracking. The RPFSBLS cascades three components: a Positive Recurrent Polynomial Fuzzy Neural Network (PRPFNN) with input positivity normalization and a bounded recurrent compensator, an intermediate normalization layer, and a Recurrent Stacked Broad Learning System (RSBLS) with per-block temporal self-feedback for hierarchical feature extraction. Operating as a one-step-ahead online identifier, RPFSBLS provides a real-time system Jacobian used to continuously adapt PID gains via closed-form gradient projection. A composite Lyapunov analysis formally establishes exponential convergence of the identifier error and asymptotic stability of the closed-loop tracking error under an explicitly derived adaptive learning rate. Validation on the two-tank level control system and a four-zone plastic extrusion barrel has demonstrated consistent improvements with over 8.55% RMSE and 17.13% ISE reductions on the two-tank and 7.26% RMSE and 7.83% ISE reductions on the extrusion barrel over PID.

Keywords: Control of Uncertain Systems, AI and Applications, Intelligent Industrial Systems
Abstract: This paper proposes an adaptive predictive PID control (APPID) strategy utilizing an interval Type-3 fuzzy broad learning system with elliptic membership functions (IT3-EFBLS) for setpoint tracking of nonlinear discrete-time dynamic systems and industrial plants. Three theorems are established to find sufficient conditions of identification convergence and closed-loop stability of the overall IT3-EFBLS-APPID controller. The comparative simulation studies provide a comprehensive evaluation of the computational complexity and control efficacy of the proposed approach relative to existing methods. Performance comparisons are rigorously assessed in terms of setpoint tracking, disturbance attenuation, and robustness against model uncertainties. Particular effort is paid to conduct a quantitative assessment of the computational overhead per sampling instance, specifically adhering to strict real-time deterministic constraints. Simulation results verify the superior effectiveness of the IT3-EFBLS-APPID in controlling complex nonlinear dynamic systems with significant uncertainties. Experimental validations conducted on the industrial vacuum heating oven are provided to demonstrate the practical feasibility of the proposed method in high-precision industrial applications.

Keywords: Reinforcement Learning, Transfer Learning, Agent-Based Modeling
Abstract: Pedestrian simulation is widely used in evacuation analysis, architectural design, and virtual environments. Although reinforcement learning (RL) enables pedestrian agents to acquire adaptive behaviors from interaction data, training from scratch in each environment is computationally expensive and often yields environment-specific policies. This paper presents a two-stage transfer-learning framework toward efficient multi-agent pedestrian learning via single-agent pretraining. In the first stage, a shared Q-network is pretrained over multiple single-agent source tasks by combining Conservative Q-Learning (CQL) with adaptive source-task sampling. In the second stage, the pretrained Q-network is used to initialize agents in a multi-agent target task and is further optimized using Deep Q-Network (DQN). Experiments on four target tasks show that the proposed initialization improves early-stage learning performance in intersection, bi-flow, and bi-door environments and significantly reduces training time in these tasks. In a bottleneck environment, clear early-stage gains are limited, but no harmful negative transfer is observed under the present experimental settings. These results suggest that single-agent pretraining provides a practical step toward improving the efficiency of multi-agent pedestrian learning.

Keywords: Kansei (sense, emotion) Engineering, Machine Learning, Image Processing and Pattern Recognition
Abstract: In recent years, communication via social media has become increasingly active; therefore, many users share daily experiences and opinions on major platforms such as X (Twitter), Instagram, and Facebook. Reputation on social media has become a more critical factor for marketing than conventional media like newspapers, magazines, and television. Posts conveying the appeal and evaluations of food significantly impact consumer purchasing behavior. Food-related posts that appeal to consumers' emotions often include ``Sizzle'' representation. Analyzing these expressions allows for a deeper understanding of shifting consumer preferences and trends, enabling more effective product development and promotional strategies. In this paper, we propose a novel method for effectively analyzing food-related text data posted on social media. We categorize a vast amount of posts based on whether they include ``Sizzle'' representation and whether they contain positive or negative opinions. For the classification task, we construct deep learning-based classification models. Furthermore, to ensure the diversity of the dataset and improve classification performance, we utilize a specialized data augmentation method based on ``Sizzle'' representation. We conducted experiments using posts from X. The experimental results showed that BERT combined with this specialized data augmentation achieved the highest classification accuracy in our evaluation.

Keywords: Generative AI, AI and Applications
Abstract: With the increasing demand for AIGT detection in recent years, evasion techniques have gradually become a prominent research focus. AI-generated text (AIGT) detection evasion aims to reduce the detection probability of AIGT, helping to identify weaknesses in detectors and enhance their effectiveness and reliability in practical applications. Previous evasion methods have relied on manually crafted modification strategies, such as the selection of replacement words and hand-designed examples, which require expert domain knowledge. To address the above limitations, we propose the Tree of Evading Detection (ToED). ToED employs a two-tier mixed prompt to construct a tree structure that guides Large Language Models (LLMs) in autonomously exploring optimal modification strategies, thereby enhancing the ability of AIGT to evade detection. Experimental results demonstrate that our method effectively reduces the average detection accuracy of various AIGT detectors across texts generated by different LLMs, surpassing that of two other baselines and achieving the best performance in evading detection.

Keywords: Generative AI, Knowledge Acquisition, Human-Machine Cooperation and Systems
Abstract: Reducing hallucinations in Large Language Models (LLMs) is essential for accurate data extraction from large text corpora. Current methods, like prompt engineering and chain-of-thought prompting, focus on individual documents and fail to consider relationships across a corpus. This paper introduces Peer Context Outlier Detection (P-COD), which uses inter-document relationships to improve extraction accuracy in scientific literature summarization, where papers with similar experiment settings should draw similar conclusions. By comparing extracted data to validated peer information within the corpus, we adjust confidence scores and flag low-confidence results for expert review. Our experiments demonstrate up to 98% precision in outlier detection across 6 scientific domains, reducing hallucinations and letting researchers focus on genuinely ambiguous cases.

Keywords: Representation Learning, Generative AI, Autonomous Vehicle
Abstract: Reconstructing dense 3D scenes from sparse LiDAR point clouds (LiDAR scene completion) is a fundamental challenge in autonomous driving, where diffusion models offer a promising solution. However, existing approaches rely on object-level autoencoders that collapse into unstable global representations at outdoor scale, and suffer from ground truth data corrupted by odometry drift that systematically degrades supervision quality. Furthermore, multi-step diffusion inference incurs prohibitive latency for real-time deployment. We present CloudDiffusion, addressing these issues with three independent components. First, a multi-token Gaussian VAE with cross-attention pooling provides stable scene-scale LiDAR compression as a standalone reconstruction module, avoiding the global-pooling and codebook-collapse failure modes of prior point-cloud autoencoders. Second, an anchor-based ICP ground truth refinement pipeline eliminates drift-induced noise from training supervision, reducing our single-step x0 diffusion teacher's squared Chamfer distance by approximately 16x on SemanticKITTI seq. 08 (0.396 to 0.024 m^2) with no model change (partly aided by the denser, more compact refined references). Third, the same teacher completes scenes in a single x0 step, operating directly in coordinate space, not in the VAE latent. It runs in near real time at 209ms/frame, 65-138x lower inference latency than iterative diffusion baselines. Our results indicate that data quality dominates model design in this regime, and suggest that multi-token latent spaces could serve as a stable first stage for future latent diffusion-based scene completion.

Keywords: Human Performance Modeling, Human-Computer Interaction, Human-Machine Cooperation and Systems
Abstract: Dance imitation integrates motor planning, sensorimotor integration, and social cognition, offering a sensitive framework to characterize motor behavior in autism. In this work, we explore a computational analysis framework to identify potential biomarkers that allow the design and development of improved medical and human-machine systems. We analyzed 3D motion capture data from autistic and neurotypical adults performing dance imitation under solo and socially-framed duo conditions. Methodologically, using Dynamic Time Warping, we quantified movement consistency and propose the Social Context Sensitivity Index (SCSI) to measure modulation of variability by social framing. These features were then used on a classifier to discriminate subjects into autistic or neurotypical groups. Results show that neurotypical adults exhibited increased movement variability in socially-framed imitation, especially in upper and lower limbs, whereas autistic adults maintained consistent movement across contexts. Classification achieved 79.2% balanced accuracy in distinguishing groups. These findings suggest that social context sensitivity in motor imitation constitutes a robust biomarker of autism-related motor behavior, highlighting the importance of social modulation in motor assessments and informing the development of inclusive human-centric technologies.

Keywords: Human-Computer Interaction, Virtual, Augmented, Mixed Reality, Kansei (sense, emotion) Engineering
Abstract: In virtual reality (VR) services, users' avatars naturally approach each other, freely create conversation spaces, and engage in conversations. It is thought that users may infer the conversational atmosphere of a group by observing avatar behavior, such as mouth movements, frequency of overlap, and gaze direction, before joining the group. However, it remains unclear which specific avatar behaviors reflect these inferences of conversational atmosphere, and current VR services do not account for how avatar behaviors influence perceptions of conversational atmosphere. If effective behaviors that facilitate joining a conversation are identified, users could more easily decide which conversation group to join. This study examined turn-taking behaviors (overlap, gap, and gaze rules) in conversations and investigated which behaviors effectively support joining conversations in VR. In addition, it was assumed that the user's avatar would be positioned some distance away from the conversation group, and the user could not hear the group's voices because they had not yet joined the conversation. Several types of conversations were prepared based on different avatar behaviors. Participants completed paired comparisons, indicating which conversation group they would prefer to join. The experimental results demonstrated the significance of gaze rules in judging conversational atmospheres and showed that mouth movements alone were insufficient for this purpose. These findings contribute to implementing effective behavior that can create conversational atmospheres in VR environments. They are also useful for designing conversational artificial intelligence and training in VR for individuals who find it challenging to join conversation groups.

Keywords: Human-Machine Interaction, Neural Networks and their Applications, Application of Artificial Intelligence
Abstract: Surface electromyography (sEMG)-based gesture recognition enables natural human-machine interface (HMI), yet non-stationary transients in dynamic gestures pose significant challenges for real-time performance. This paper proposes a robust framework for dynamic sEMG gesture recognition using a portable wireless acquisition device. To handle signal variability, continuous data are processed through a pipeline featuring data augmentation and window-level normalization. A lightweight Real-Time Deep Temporal Architecture (RT-DTA, ~100k parameters) is employed to capture temporal dynamics, while a confidence-gated sliding voting mechanism ensures stable online predictions. Evaluating the system on 15 gestures, we achieved an offline trial-level accuracy of 96.67% ± 5.04% and an online accuracy of 95.83% ± 4.35%. Crucially, the system achieved a median command output time of 0.51 s, demonstrating its ability to decode motor intent during biomechanical transitions. These results confirm the framework's high reliability and low computational cost for practical, multi-gesture HMI applications. The source code and data are available at: https://doi.org/10.6084/m9.figshare.26764714.

Keywords: Human-Machine Cooperation and Systems, Systems Safety and Security, Cooperative Systems and Control
Abstract: Autonomous AI agents, increasingly empowered by large language models, are becoming important components of human-machine systems for high-stakes decision support in digital twin ecosystems. However, existing multi-agent systems often lack robust verification for identity, capability, and policy compliance, especially in decentralized environments spanning multiple institutions. This paper proposes a neuro-symbolic decentralized governance framework for verifiable agents in collaborative digital twin environments. By representing agents through multi-layer semantic profiles, the framework bridges probabilistic neural reasoning with deterministic institutional governance, thereby supporting trustworthy human-AI collaboration and meaningful human oversight. Capabilities are grounded in formal domain ontologies to enable machine-interpretable, policy-aware, and context-sensitive participation. These credentials, issued by organizational authorities, are validated via blockchain-based smart contracts, ensuring auditable participation without exposing sensitive data. We demonstrate the framework using a decision-support prototype with clinic, digital twin, and wearable provider agents effectively prevents unauthorized interaction and enforces institutional policies with manageable overhead. Our findings suggest that neuro-symbolic decentralized governance provides a scalable and trustworthy pathway for safe human-machine collaboration across institutional boundaries.

Keywords: Human Performance Modeling, Representation Learning, Human-Machine Interaction
Abstract: Aging workforces and labor shortages in the construction industry have created a growing demand for technologies that improve productivity and operational safety. Achieving this goal requires an understanding of operator states, including skill level and workload, which in turn calls for methods that can objectively characterize operational patterns from machinery data. In this paper, we apply a variational autoencoder (VAE) to hydraulic excavator operation data and propose a latent-space analysis method that captures operational patterns across multiple temporal scales. Multi-scale analysis with short- and long-time windows enables simultaneous characterization of instantaneous operational features and work-process-level consistency. Experiments with nine operators show that skilled operators exhibit higher operational consistency, while some unskilled operators show a progressive decline across cycles. In binary classification using latent-space metrics, the best AUC-ROC (0.822) was achieved under the two-dimensional multi-scale condition, outperforming either scale alone. Furthermore, significant correlations with bucket-tip-trajectory metrics suggest that the latent representations encode physically grounded motion structure. These results indicate that VAE-based latent-space analysis effectively captures skill-related operational characteristics and can serve as a promising foundation for operator state monitoring in construction machinery.

Keywords: Human-Machine Interaction, Human Enhancements, Affective Computing
Abstract: Body extensions have primarily been developed to augment physical capabilities, but their use for regulating internal states, such as inducing relaxation, remains underexplored. This study focuses on using a tail to regulate the user’s internal state without interfering with primary tasks. The tail can be moved either actively or passively and provides vibrotactile feedback to the user’s back. We investigated the effects of the tail on users' internal states during a tension-inducing announcement task in virtual reality. Subjective arousal increased from pre- to post-task in all conditions and was higher after the task in active control than in passive control and the no-tail condition. During the task, heart rate was lower in passive control than in the no-tail condition, whereas active control elicited greater agency and effort than passive control. These findings suggest a trade-off between agency and operational demand and indicate that passive rhythmic tail feedback may attenuate task-related physiological arousal.

Keywords: Human Factors
Abstract: Team progress, a sequence of directional state transitions at the utterance level, captures the human factors of how teams dynamically advance toward goals. However, current studies focus on retrospective survey-based measurements over a period of time, which ignore the dynamic utterance-level directional changes that accumulate to influence team performance. Furthermore, current coding methods mainly rely on expert manual labeling, which is labor-intensive and limits their potential for timely feedback. To address these limitations, this study proposes a framework for human-centered communication progress monitoring in human system interaction simulation. Using LLM-based augmentation and context-aware modeling, our approach achieves 0.94 accuracy with a lightweight shared subspace bi-linear MLP. Ablation shows that short-horizon memory (k=3) is sufficient for optimal prediction.

Keywords: Human-Machine Cooperation and Systems, Human-Computer Interaction, Intelligent Industrial Systems
Abstract: Multidisciplinary engineering teams in safety-sensitive process industries face persistent collaboration challenges that affect project quality, efficiency, and schedule adherence. Although AI offers increasing potential to support engineering work, there is limited empirical knowledge about what practitioners require from AI-enabled systems in real engineering settings. We conducted semi-structured interviews with 20 practitioners and applied thematic analysis to identify four challenge dimensions: structural communication barriers, technology gaps, process misalignment, and domain knowledge disparities. Based on these findings, we derived 13 functional, 8 non-functional, and 7 organizational requirements. These requirements emphasize integration with existing discipline-specific tools rather than the replacement of established engineering platforms. We also propose a conceptual integration-layer multi-agent reference architecture that translates the requirements into design guidance for human-in-the-loop decision support. The findings show that trust, explainability, auditability, and subject matter expert validation are important conditions for AI adoption in safety-sensitive engineering environments. Overall, this work provides empirically grounded requirements and design guidance for aligning AI capabilities with the socio-technical realities of multidisciplinary engineering collaboration.

Keywords: Human-Collaborative Robotics, Haptic Systems, Human-Machine Interaction
Abstract: Teleoperated laparoscopic systems often use articulated robotic tools controlled to pass through a remote centre of motion (RCM) at the entry point, restoring two degrees of freedom (DOF) lost during manual laparoscopy. When standard, non-articulated instruments are used, the RCM introduces kinematic redundancy between the human operator and the robot, as the console interface retains more DOF than the task requires. Existing approaches resolve this redundancy at the robot or task level, but do not explicitly address how it should be handled at the human interface. This paper treats redundancy as a user interaction problem and proposes two distinct strategies to enforce the RCM directly at the interface. In the first, the user is free to translate the device while its pitch and yaw are actively controlled to align with the tool axis. In the second, the user is free to orient the haptic device handle, and its position is regulated through force feedback to remain aligned with the RCM. Both methods are implemented and evaluated in a teleoperated simulation with three user groups performing identical tasks. The results show that both haptic strategies improve task efficiency and motion quality compared to no feedback. While overall performance is comparable between methods, the force-based approach better enforces alignment with the RCM, consistent with the higher force output capability of the device. These findings suggest that redundancy can be leveraged to enable different modes of user interaction with the environment, while maximizing the capabilities of the haptic interface without compromising intuitiveness or performance.

Keywords: Human-Machine Interaction, Human-Machine Cooperation and Systems, Artificial Social Intelligence
Abstract: Recent advances in large language models have enabled dialogue agents to generate fluent and human-like responses by incorporating persona modeling and long-term memory. However, most existing systems rely on a fixed response selection strategy, implicitly assuming that a single configuration is effective for all users. Such one-size-fits-all approaches are misaligned with the goal of human-centric intelligence, which requires adaptation to individual differences. This paper proposes a personality-aware strategy optimization framework for persuasive dialogue, where response selection is formulated as a multi-objective decision process balancing goal achievement, contextual coherence, and personality adaptation based on the Big Five model. Through controlled experiments on a persuasive donation task, we show that while persuasion success often reaches ceiling levels, dialogue efficiency—measured by the number of turns required to reach agreement—varies substantially depending on the interaction between personality traits and strategy weightings. Statistical analyses reveal significant interaction effects, indicating that optimal strategies differ across personality profiles. These results highlight dialogue efficiency as a human-centric performance indicator that may serve as a plausible proxy for user burden and cognitive effort in human–agent interaction. In addition, a questionnaire study demonstrates a certain degree of alignment between LLM-based response evaluations and human judgments, while revealing trait-dependent differences in evaluation tendencies in isolated response selection, with LLM-based simulation serving as a controlled testbed rather than a proxy for absolute human persuasion performance.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Machine Vision
Abstract: Frequency-domain image restoration typically assumes that noise and signal are spectrally separable. However, in outdoor autonomous vision, adherent raindrops exhibit sharp boundaries and irregular geometries, resulting in a broad spectral distribution that undermines frequency-based filtering. In this study, we investigate both the limitations of global frequency filtering and an alternative use of frequency information through a dedicated loss function. We compare five segmentation architectures on the RWRSD dataset under a strictly separated scene-level cross-validation protocol. HRNet-W32, as a representative of spatially coherent architectures, achieves higher mIoU (+0.40pp) and BoundaryIoU (+0.90pp) than the FSI-augmented model at 6.2× lower cost (32.7 vs. 204.2 GFLOPs), with the most stable fold-wise rank ordering (rank std = 0.50)—evidence that spatial coherence is a more reliable prior than spectral filtering in uncontrolled environments. We further propose the Hybrid Coherence Loss (L_HC), which supervises the frequency-domain consistency of predicted segmentation masks rather than filtering the image. Ablation experiments reveal that L_HC is ineffective when applied in isolation (1/5 models improve), but becomes consistently beneficial when combined with spatial loss (4/5 models improve, up to +2.80pp BoundaryIoU for HRNet-W32). These findings suggest a design principle for raindrop segmentation in uncontrolled environments: frequency information should guide prediction coherence, not filter the signal.

Keywords: Image Processing and Pattern Recognition, Neural Networks and their Applications, Deep Learning
Abstract: Deepfake detection systems often exhibit significant performance degradation when deployed on unseen manipulation methods, limiting their reliability in real-world multimedia environments. This lack of generalization poses critical challenges for misinformation mitigation, digital forensics, and human-centric AI systems. Existing detectors perform well on the forgery methods they are trained on, but their accuracy drops sharply on unseen pipelines. To bridge this generalization gap, we propose WAVIE (Wavelet Augmented Vision Intermediate Embeddings), an end-to-end architecture that combines complementary spatial and frequency cues on top of a frozen CLIP backbone. WAVIE projects intermediate transformer embeddings through a lightweight learnable module, applies a three-level Daubechies-6 (db6) discrete wavelet transform (DWT), refines the low-frequency branch while preserving the high-frequency branch, reconstructs the feature via inverse DWT, and performs classification. Trained only on FaceForensics++, WAVIE achieves AUROC = 0.852 on Celeb-DF-v1 [3], 0.852 on Celeb-DF-v2 and 0.831 on WildDeepfake at the frame level, achieving superior state-of-the-art Generalization baselines. Extensive ablation studies confirm the importance of both the wavelet module and the intermediate-feature aggregation for cross-dataset performance, highlighting the necessity of jointly leveraging spatial and frequency domains. These results position WAVIE as a strong baseline for deepfake detection in the wild.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Wearable Computing
Abstract: Edge video compression for surveillance requires balancing bitrate, reconstruction quality, computational cost, and energy consumption on resource-constrained edge devices. This paper presents an asymmetric video compression frame-work consisting of a lightweight spiking neural network (SNN) encoder and a convolutional neural network (CNN) decoder for fixed-camera grayscale video. Each 8-bit frame is decomposed into eight parallel bit planes, while the encoder generates a binary spike latent that is bit-packed and losslessly compressed using zlib. The training objective combines single-scale struc-tural similarity (SSIM), optical-flow-guided motion loss, and firing-rate regularization to control coding rate. Using a com-mon non-overlapping test set containing 9,209 frames, we com-pare the proposed method with MCUCoder, H.264, and H.265. The conventional-codec references and MCUCoder provide stronger rate-distortion performance over the evaluated ranges. However, for the current Edge Encoding implementations, which include preprocessing, transfer, encoder inference, and entropy coding, the proposed method achieves higher through-put and 45–68% lower idle-subtracted measured-component energy than MCUCoder. An analytical event-driven hardware model separately estimates approximately 87% lower encoder energy by exploiting sparse spike activity. This estimate rep-resents the potential of dedicated hardware that skips inactive synaptic operations, not a measurement of the current dense float32 GPU implementation. The results identify the proposed architecture as a promising event-driven edge video encoding framework, while its practical hardware efficiency remains to be demonstrated on a common embedded target.

Keywords: Image Processing and Pattern Recognition, AI and Applications, Computational Life Science
Abstract: Research on nematodes is important because they serve as simple, well-understood model organisms that help scientists uncover fundamental biological processes relevant to human development, genetics, and disease. We present a machine learning model to distinguish strains of Caenorhabditis elegans (C. elegans) from single worm videos. We considered 16 features related to motion and body characteristics, using four statistical measurements across frames with state-of-the-art classifiers (Multilayer Perceptron, Support Vector Machine, Extreme Learning Machine, Bayesian Linear Discriminant Analysis, and K-Nearest Neighbors). Feature selection was assessed using forward selection and backward elimination. The results support the conclusion that dividing videos into smaller clips yields a 9.34% increase in mean accuracy, and using additional statistical measurements beyond the mean yields a 4.33% increase in mean accuracy. The results also indicate that postural features, such as bend length standard deviation and tortuosity, have the highest discriminant power.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Image super-resolution (SR) aims to reconstruct high-quality images from degraded low-resolution inputs. Transformer-based methods have shown strong performance by modeling long-range dependencies; however, many existing approaches introduce redundant attention computations or rely on fixed interaction patterns, leading to suboptimal efficiency. To address these limitations, we propose ELMAN, an Efficient Long-range Multi-head Attention Network for image SR. ELMAN employs shift convolution to capture local structural information and introduces a Group-wise Multi-scale Multi-head Self-Attention mechanism, which decomposes attention across non-overlapping feature groups with shared projections to reduce redundancy while preserving global context modeling. These components are integrated into an Efficient Long-range Multi-head Attention Block. In addition, a shared attention strategy is adopted to further reduce computational complexity. Extensive experiments on standard benchmark datasets demonstrate that ELMAN achieves a favorable trade-off between reconstruction accuracy and computational efficiency, delivering competitive performance with reduced FLOPs.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Open-vocabulary object detection aims to localize and classify objects beyond the fixed set of categories seen during training. Recent open-vocabulary object detection methods improve localization and classification for unseen categories by leveraging frozen VLM as a detector backbone. However, VLM classification score lacks recognizing position and scale of the object in an image. We observe that pretrained VLMs enable to classify foreground and background regions. According to this observation, we propose a simple inference-time Proposal Calibration (ProCal) that improves localization quality of the classification score. ProCal computes a proposal prior by combining two scores: localization-aware foreground score and background-aware suppression score. Localization-aware foreground score captures whether a proposal contains an object area. Background-aware suppression score measures the extent to which the proposal resembles background. We analyze that ProCal suppresses false novel activation on background proposals and consistently ranks true novel proposals above background and partial novel proposals. Applied to CLIPSelf ViT-L/14, ProCal improves mask APr +2.5 on the OV-LVIS. The analyses show that proposal-level localization-aware reranking effects to mitigate ranking miscalibration for novel objects.

Keywords: Image Processing and Pattern Recognition, Representation Learning, Deep Learning
Abstract: Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image from a multimodal query consisting of a reference image and an edit text describing the desired modification. Recent ZS-CIR studies have relied on projection-based methods that map reference image into pseudo-word tokens in the text embedding space. However, such studies require additional projection and re-encoding steps, increasing training complexity, reducing efficiency, and introducing a discrepancy between training and inference. In this paper, we propose DiCE-CIR, a direct composition learning method that predicts a composed query representations by directly composing a reference image and an edit text. To enable scalable training without manually annotated triplets, we automatically construct compositional training samples from large-scale image-caption pairs using a large language model to generate edit texts and target captions. Based on these samples, we train a lightweight composition module with objectives that promote alignment with the target, edit-consistent semantic transformation, and retrieval discriminability. We conduct extensive experiments on ZS-CIR benchmarks and show that DiCE-CIR achieves state-of-the-art performance on CIRCO and competitive performance on CIRR while maintaining high computational efficiency.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Machine Learning
Abstract: In recent years, the rapid growth of 5G communication networks and IoT devices has significantly increased the demand for real-time video transmission. Meanwhile, the rise in video resolution and frame rates has led to increased data volume and computational requirements, making efficient video compression essential under limited bandwidth and hardware resources. Conventional standards such as H.264/AVC and H.265/HEVC achieve high compression efficiency, but at the cost of increased computational complexity, power consumption, and hardware cost.To address these challenges, this paper proposes an FPGA-based spiking neural network (SNN) autoencoder for energy-efficient video processing in edge AI systems. The proposed method leverages sparse spike representations, event-driven computation, and a multiplier-free architecture to significantly reduce computational complexity and hardware resource usage. In addition, a bit-plane decomposition-based input scheme enables high-throughput and low-latency processing.Implemented on a Kria KR260 FPGA, the proposed system achieves competitive reconstruction performance while significantly reducing power consumption and DSP utilization compared to ANN-based methods and H.264. These results demonstrate the effectiveness of SNNs for low-power, real-time video processing on edge devices.

Keywords: Image Processing and Pattern Recognition, Neural Networks and their Applications
Abstract: Remote identity authentification using Identification Documents has been a major challenge for several years. DeepFakes advent and the development of AI-guided tools helps fraudsters creating counterfeit ID Documents. Ensuring the authenticity of ID Documents has become a real clue in the seurization of remote authentification. This need is all the more pressing given the increasing digitization of administrative and transactional processes. To ensure widespread accessibility, the system should rely solely on video captured via mobile devices. In this specific context, confirming the authenticity of ID is a real challenge as many security features needs specific device like infrared sensor for instance. Among underutilized but promising security features, holographic printings hold a special place. Difficult to counterfeit, they produce distinctive visual effects according enlightment, making them both detectable in a video captured by a smartphone camera and difficult to imitate. In this paper, we propose a Remote Identity Document Verification System (RIDVS) and an approach based on a video transformer for detecting holograms in simple videos captured by smartphones. Our system is designed for a smartphone-based capture process, followed by a server-side verification. The hologram detection method builds on a robust model previously validated in a related research domain. We demonstrate that it outperforms existing SotA methods, achieving near-perfect accuracy even when trained on medium- to small-sized datasets. In particular, we report improvements of +26.86% in Recall and +17.93% in accuracy over the best MIDV-Holo baseline. This study includes several experiments that evaluate the model adaptation to frugality, both for training samples and computational resources.

Keywords: Adaptive Systems, Optimization Control, Intelligent Transportation Systems
Abstract: The control of autonomous riderless bicycles remains a challenging problem in nonlinear control research. Traditional model-based controllers, such as Feedback Linearization (FL), can achieve stabilization but rely on simplified dynamic models that may not capture unmodeled or time-varying effects. Recent data-driven control methods, including Data-enabled Predictive Control (DeePC) and Data-enabled Policy Optimization (DeePO), allow closed-loop optimal control directly from measured input-output data, bypassing explicit system identification. This paper presents a cascaded control architecture that integrates FL, DeePO, and DeePC for simultaneous stabilization, velocity regulation, and trajectory tracking of an autonomous bicycle. First, a multi-input-multi-output (MIMO) FL controller is derived to regulate the lean angle and forward velocity. Then, DeePO is employed to adaptively enhance the FL control law using closed-loop data generated under persistently exciting inputs. As a supplementary contribution, a new forgetting-factor integration scheme was introduced into the standard DeePO algorithm to enhance its performance under nonlinear dynamics and time-varying system parameters. Finally, DeePC operates in the outer control loop, generating reference trajectories for the FL+DeePO controller via regularized data-driven optimization. The proposed cascaded FL-DeePO-DeePC approach is validated through high-fidelity simulations in Simscape.

Keywords: Human-Centered Transportation, Intelligent Transportation Systems, Smart Buildings, Smart Cities and Infrastructures
Abstract: As automated vehicle technologies advance, improving passenger comfort during braking remains a key challenge. This study presents an experimental analysis of human braking behavior using data collected from controlled vehicle tests at the Virginia Tech Transportation Institute (VTTI), including synchronized measurements of velocity, acceleration, braking torque, and remaining stopping distance across varying speeds and distances. The analysis reveals how braking intensity and deceleration profiles evolve as a function of speed and available stopping distance, establishing a data-driven baseline for comfort-related braking behavior. Based on these findings, a trip comfort index (TCI) is first developed to quantify braking comfort based on time spent within defined comfort zones, and an enhanced metric, TCI*, is further proposed to incorporate both deceleration intensity and comfort-zone exposure. These results provide insight into how comfort is affected under different operating conditions and offer guidance for the design of comfort-aware automated driving strategies, including the deployment of vehicle-to-infrastructure (V2I)-enabled technologies to support earlier, smoother braking.

Keywords: Intelligent Transportation Systems, Environmental Sensing, Modeling of Autonomous Systems
Abstract: LiDAR-inertial odometry (LIO) systems typically rely on IMU accelerometers for velocity estimation of internal state, which is however susceptible to IMU drift in practical irregular paths and long-term localization. The emerging frequency-modulated continuous wave (FMCW) LiDAR can directly obtain the absolute velocity from the point cloud through the Doppler effect, which has the potential for high-precision and real-time velocity estimation from single scan. However, due to the scarcity of device types and changes in modal information, publicly available datasets and algorithms for FMCW-based LIO are still in their initial stages. This paper proposes a fusion odometry of FMCW LiDAR and IMU, where the velocity observations measured by FMCW are deeply fused with the IMU pre-integration information to achieve continuous estimation of states and correction of the IMU bias. A multi-residual optimization framework is established to improve the accuracy and significantly enhance the robustness of the system. Experiments were conducted using a self-developed spinning FMCW LiDAR across several scenarios and trajectories, as well as an emerging public dataset. The results showed that even using the low-cost FMCW LiDAR, Doppler velocity estimation can markedly enhance the accuracy and stability of LIO systems, outperforming other classic open-source LIO solutions.

Keywords: Intelligent Transportation Systems, Distributed Intelligent Systems, Application of Artificial Intelligence
Abstract: Conventional air traffic control divides airspace into specific regions, creating a scaling bottleneck as traffic grows. Choosing how to partition airspace is not straightforward because grid size affects workload, handoff frequency, and the capacity of whatever coordination mechanism operates within each sector. We present a three stage pipeline that automates sectorization and sector coordination while preserving human oversight. First, a two stage XGBoost classifier predicts the optimal 3D grid configuration from 23 location-agnostic traffic features, achieving 91.38% accuracy on a 65,000 sample dataset derived from Federal Aviation Administration System Wide Information Management replays. Second, a leaderless Paxos consensus protocol lets aircraft coordinate sector entries among themselves, maintaining above 96% entry success with low near mid-air collision rates across all tested configurations. Third, Bayesian Optimization with a Gaussian Process surrogate tunes eight protocol parameters per airport in 50 trials, revealing that each traffic environment requires a qualitatively different configuration. The resulting pipeline offers a practical path toward scalable, autonomous airspace management as traffic demand outpaces controller capacity.

Keywords: Machine Learning, Wearable Computing, Human-Computer Interaction
Abstract: In electromyogram (EMG)-based motion recognition, it is impractical to predefine all motions that may be required during deployment, necessitating class-incremental learning that sequentially adds new motion classes. The primary challenges in class-incremental learning are catastrophic forgetting, where previously acquired knowledge is overwritten when learning new classes, and the memory cost of retaining past data to counteract it. In particular, for EMG-based motion recognition intended for edge devices with limited computational resources, it is essential to suppress catastrophic forgetting and maintain low memory cost. In this paper, we conducted a comparative evaluation of eight class-incremental learning methods spanning generative and discriminative approaches, including both deep and non-deep learning methods, for EMG signal classification. Using four datasets, we evaluated each method in terms of classification accuracy, backward transfer, and memory cost. The results demonstrated that deep learning-based methods suffered significant accuracy degradation from catastrophic forgetting as the number of tasks increased, whereas generative models maintained stable accuracy with low memory cost. Among generative models, the scale mixture classification model (SMCM), which captures EMG signal variability, achieved the most favorable accuracy–memory trade-off while effectively suppressing catastrophic forgetting across all datasets.

Keywords: Information Assurance and Intelligence, Machine Learning, Visual Analytics, Communication
Abstract: The proliferation of generative AI models has made it increasingly difficult to trace the origins of synthetic images, posing a significant challenge to digital media forensics. While existing methods focus primarily on binary detection of synthetic content, the task of model attribution (MA) identifying which generative model produced a given image is essential for accountability and source tracking. However, most current MA approaches rely on retraining deep models from scratch with large labeled datasets whenever new generative models emerge, making them inefficient and unscalable. In this work, we formulate model attribution as a few-shot class-incremental learning (FSCIL) problem and propose FSCI-MA, a novel framework that incrementally incorporates new generative models with minimal data and without full retraining. To address this, we exploit multilevel feature representations extracted from CLIP-ViT and introduce an Adaptive Integration Module (AIM) that computes a dynamic, learnable aggregation of intermediate block features for each input. Extensive evaluations across a wide range of generative models (including GANs and diffusion-based models) demonstrate that FSCI-MA significantly improves scalability and adaptability, outperforming baseline MA and FSCIL methods in both attribution accuracy and generalization to unseen models.

Keywords: Machine Learning
Abstract: The rapid growth of online textual data has intensified the need for machine learning systems that remain robust to dynamic, continuously shifting data distributions, a phenomenon known as concept drift. Such drift can rapidly degrade model performance, demanding timely adaptation or retraining - an especially challenging requirement in continuous text stream mining settings. To enhance robustness in the classification task of textual data streams, we propose DA-Dual‑SBERT, a drift-aware dual‑model framework built upon Sentence‑BERT. When a drift is detected, the older model, i.e., the legacy model, is fine-tuned considering recent stream samples stored in a buffer. Instead of leveraging all buffered instances, DA-Dual‑SBERT employs a WordPiece Token Ratio–based sampling strategy to prioritize the most informative instances for adaptation. Extensive experiments across multiple drifting scenarios demonstrate that DA-Dual‑SBERT consistently exhibits greater drift resilience, achieving the highest Macro F1‑scores and the best overall ranking compared with competitive baselines.

Keywords: Medical Informatics, Representation Learning, Machine Learning
Abstract: Electroencephalography (EEG) models used for epilepsy are often limited to specific datasets and tasks. This limited approach can make it challenging to apply these models across different datasets or in various situations. However, recent studies in foundation models and self-supervised learning suggest that an adaptable EEG backbone could support a range of EEG related tasks. In this study, we have developed a multimodal EEG foundation model that combines a raw signal encoder based on the Mamba architecture, a Vision Transformer (ViT)-style encoder for time-frequency data, and a lightweight encoder for text, all within a shared embedding space. The pretraining process relies on several innovative techniques, such as masked modeling, cross-view contrastive alignment, and temporal consistency losses. These methods are designed to create rich, seizure-relevant representations without requiring labeled data. To assess the efficacy and generalization of our pretrained model, we fine-tuned it on the canonical CHB-MIT seizure detection benchmark and additional seizure detection datasets, and conducted extensive experiments comparing different model variants. On the standard CHB-MIT split, our best single model achieved an AUROC of 0.874, and an ensemble variant reached 0.878 AUROC, representing state-of-the-art performance on this benchmark. In addition to standard train-test splits, we evaluated performance under a leave-one-subject-out (LOSO) protocol, which is rarely reported in prior EEG seizure modeling work and highlights the difficulty of patient-independent seizure detection, with a mean LOSO balanced accuracy of 0.558 across 19 subjects. Across datasets and evaluation settings, our multimodal foundation model enabled robust seizure detection and straightforward adaptation to new seizure detection scenarios, while also supporting interpretable seizure localization. These findings highlight the potential of EEG-specific foundation models as versatile tools that can be reused not just for seizure detection but also for broader analyses of EEG data related to seizures.

Keywords: Machine Learning, Generative AI, AI and Applications
Abstract: The rapid proliferation of AI-generated content raises a critical systemic risk: as synthetic data increasingly contaminates future training corpora, recursively trained models may suffer progressive performance degradation—a phenomenon termed model collapse. While prior work has demonstrated the existence of this phenomenon, principled mitigation strategies with quantifiable safety thresholds remain largely unexplored. In this paper, we formalize recursive AI training as a dynamic system and empirically quantify the conditions under which model collapse occurs. We propose Residual Human Data Injection (RHDI), a simple yet effective mitigation strategy that preserves a fixed anchor ratio ƒ¿ of authentic human data in each training generation. Through controlled experiments on the IMDb sentiment classification benchmark using a GPT-2- based synthetic data generator over 20 recursive generations, we find that without any human data anchor (ƒ¿ = 0), model accuracy degrades from 83.6% to 62.1% (a 25.7% relative drop) and vocabulary diversity collapses irreversibly. In contrast, RHDI with ƒ¿ ≥ 0.5 completely prevents performance collapse across all 20 generations, maintaining accuracy within 3.7% of the humandata oracle. Our results establish ƒ¿ ∗ = 0.5 as a practical safety threshold and provide the first systematic quantification of the relationship between anchor ratio and collapse resistance. Index Terms—model collapse, recursive training, synthetic data, data poisoning, residual human data injection, AI reliability, cybernetics

Keywords: Machine Learning
Abstract: Fuel demand forecasting is difficult due to the structural heterogeneity, volatility, and region-specific consumption patterns of energy markets. Time series foundation models (FMs) show strong zero-shot generalization, but their performance varies across datasets and temporal structures, and no single FM dominates all scenarios. We propose a Mixture-of-Experts–inspired routing mechanism that dynamically selects and combines multiple zero-shot FMs. A lightweight router, trained while all experts remain frozen, assigns instance-specific weights to TimesFM, Moirai, Chronos, Timer, and Time-MoE, enabling adaptive expert selection without compromising the models' zero-shot nature. We evaluate FM-MoE on 216 monthly time series of Brazilian petroleum derivative sales spanning 34 years, where it achieves the lowest average RMSE across all forecast horizons. This suggests that sparse, lightweight routing trained on domain-specific data is an effective strategy for combining the complementary strengths of pre-trained models in zero-shot fuel demand forecasting.

Keywords: Fuzzy Systems and their applications, Machine Learning, Knowledge Acquisition
Abstract: Fuzzy inference is a general form of triangular fuzzy reasoning, that has proven to be effective at solving various types of inference problems. In this paper, we frame the problem of parameter adjustment for fuzzy rules not as a tuning problem but rather, as a design problem. We focus in particular on how to set initial values, and how to determine the learning sequence of consequent and antecedent parts. We discuss the accuracy of the proposed design methods in light of quantitative evaluations performed on sample datasets.

Keywords: Neural Networks and their Applications, Sustainable Systems
Abstract: Spiking Neural Networks (SNNs) have been used successfully in developing novel event-driven machine learning approaches. Brains combine the activity of spiking neurons with the energy-management function of capillaries and astrocytes. Biological systems have been under evolutionary pressure for millions of years to use the limited energy resources efficiently. The combination of fast spiking and relatively slow metabolic dynamics is a potential research direction to further improve energy efficiency in artificial designs, which is the objective of this work. It is shown that the dynamics of SNNs can exhibit complex dynamical behavior with hysteresis due to the impact of the metabolic system. The feasibility of using the observed effects for brain-inspired dynamical memories is demonstrated. Potential benefits for the future development of neuromorphic designs are outlined.

Keywords: Social simulations, Interactive and Digital Media, Media Computing
Abstract: More than 5.66 billion people actively use social media. With this engagement, hate speech, toxicity, spam, and cyberbullying have become prevalent on these platforms, resulting in user mental health struggles and societal fear. In this paper, we propose an interpretable and intelligent framework for sentiment analysis, toxicity detection, and account authenticity classification for controlling harmful social media comments on Instagram. A pre-trained Large Language Model (LLM) is used for three-class sentiment analysis on Instagram comments. Next, the accounts posting negative comments are classified into Bot, Scam, Real, or Spam using a machine learning (ML) model trained on a fake account detection dataset and the accounts' metadata. Finally, we validate the results using an XAI framework to ensure interpretability. Results indicate that our approach achieves accuracy, precision, recall, and F1 scores of 96.90%, 96.90%, 96.90%, and 96.90%, respectively, for fake account detection from negative commentary on Instagram. Additionally, XAI analysis reveals that our XGBoost model relies on semantically meaningful, domain- relevant profile features rather than spurious correlations, and aligns closely with its global feature importance rankings: Followers: 37.9%; Posts: 30.4%; Following: 10.3%.

Keywords: Computational Intelligence, Metaheuristic Algorithms, Machine Learning
Abstract: Exploratory Landscape Analysis (ELA) provides sampling-based features to describe landscape structure and has been widely studied for invariance properties. Recently, fitness value histograms were proposed as simple high-level features, showing strong ability to distinguish optimization problems, but their robustness to problem transformations remains insufficiently explored. This paper investigates the invariance of fitness histograms under problem transformations. The study is conducted in the context of black-box optimization using the BBOB benchmark suite of 24 noiseless single-objective functions, evaluated across multiple dimensions and transformed instances. Consistency of fitness value distributions across instances of the same problem is assessed using statistical analysis and k-Nearest Neighbors cross-validation. The results provide insight into the stability and representativeness of fitness histograms as landscape features for problem analysis.

Keywords: Machine Vision, Representation Learning, Deep Learning
Abstract: Referring Expression Comprehension (REC) localizes objects from natural-language queries. While large Multimodal Large Language Models (MLLMs) achieve strong performance, their computational cost hinders real-time deployment. Lightweight one-stage pipelines are efficient but struggle in visually cluttered scenes requiring fine-grained spatial discrimination. To address this, we propose an efficient spatial-semantic architecture with two complementary modules: a Semantic-Conditioned Feature Modulator (SCFM) for early linguistic guidance over multi-scale visual features, and a Coordinate-Aware Regression Head (CARH) for coordinate-aware localization refinement. We also provide attention visualization and diagnostic analysis of remaining failure modes. Experiments on RefCOCO, RefCOCO+, and RefCOCOg show consistent improvements on challenging splits with only minor computational overhead, preserving real-time inference. Code is available at: https://github.com/nutcliu2507/SCM-CAR.

Keywords: Machine Vision, Human-Collaborative Robotics, Cloud, IoT, and Robotics Integration
Abstract: This study focuses on robotic grasping, an essential capability for embodied intelligence and autonomous manipulation. We propose a robotic grasp detection and execution framework using the two-stage object detection model Faster R-CNN. The framework is designed to detect graspable regions for a wide range of commonplace objects. This approach was evaluated in the ROS 2 environment with Gazebo-based simulated experiments and with real-world experiments on the UR5e robotic arm to properly report the effectiveness of the framework. The model with the deepest backbone achieved an accuracy of 98.76% and 90.6% on the Cornell Grasping dataset and the Jacquard dataset with inference speeds of 32 ms. In addition, a pruned variant of the adapted Faster R-CNN architecture was developed to reduce the inference time to 7.66 ms and improve suitability for real-time robotic applications. These results demonstrate the effectiveness of the proposed framework for robust and efficient robotic grasping in both simulated and physical environments.

Keywords: Biometrics and Applications, Image Processing and Pattern Recognition, Machine Vision
Abstract: Face de-identification requires suppressing biometric identity while preserving the visual and semantic fidelity necessary for downstream analysis — a dual objective that existing generative approaches fail to satisfy due to their high-capacity latent spaces, which remain susceptible to identity leakage. We propose Latent Compression De-Identification (LCD-DeID), a structured framework that constrains identity information flow through architectural design, complementing output-level de-identification objectives. LCD-DeID coordinates three complementary mechanisms: (i) low-dimensional latent compression to structurally bound identity capacity, (ii) explicit subspace decomposition into identity-sensitive and utility-preserving components, and (iii) selective reconstruction conditioned solely on utility features. LCD-DeID achieves significant reductions in verification accuracy and embedding similarity while maintaining high structural fidelity, perceptual quality, and semantic attribute preservation. LCD-DeID establishes a deterministic, interpretable, and practically deployable baseline for privacy-preserving face synthesis in visual computing.

Keywords: Image Processing and Pattern Recognition, Intelligent Industrial Systems, Machine Vision
Abstract: Change detection models have achieved remarkable success in remote sensing applications, where they identify structural and land-use changes across temporal image pairs. However, the nature of changes relevant to industrial inspection differs substantially from those typically addressed in remote sensing: printing inspection, for instance, requires detecting diverse defect types—including subtle color shifts, texture irregularities, and fine structural deviations—that vary widely in contrast and visual characteristics. Despite growing interest in applying change detection techniques to manufacturing domains, benchmark datasets that capture these domain-specific characteristics remain underexplored. In this paper, we introduce Taktpixel2025PD-CD, a benchmark dataset for change segmentation of printing defects, designed to support per-category binary evaluation of change detection models. The dataset consists of master-target image pairs derived from actual offset and inkjet prints, each accompanied by a manually verified defect region map across six defect categories. We benchmark nine representative change detection models—spanning CNN-based, Transformer-based, and foundation-model-based architectures—under a per-category binary segmentation protocol. Our category-wise analysis reveals a pronounced performance disparity: high-contrast defects such as black spots and hairs achieve detection accuracy comparable to established remote sensing benchmarks, whereas low-contrast anomalies such as color shifts and fine linear artifacts remain substantially below practical requirements across all evaluated models.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Image inpainting seeks to restore missing or corrupted regions with visually and semantically consistent content. While recent transformer-based approaches improve global context modeling, they often suffer from high computational cost and limited efficiency. We propose a hybrid image inpainting model that combines a guided stripe Transformer with a CNN backbone for efficient global context modeling. To better restore structure and texture, the model incorporates a Masked Adaptive Feature Fusion module for multi-scale aggregation under spatial guidance, a Semantic–Texture Interaction Attention module for feature-level fusion of semantic and texture information, and an HSV-Fusion loss to enhance perceptual consistency. Together, these components enable complementary handling of global dependencies, structural guidance, and perceptual fidelity. Experiments on Places2 and CelebA-HQ demonstrate that our method achieves competitive results with high efficiency and visual quality. The source code is available at https://github.com/bunn0120/gst-net.

Keywords: Deep Learning, Machine Vision, Image Processing and Pattern Recognition
Abstract: Human pose estimation has achieved significant progress with transformer-based architectures that effectively capture long-range spatial dependencies. However, most existing approaches operate predominantly in the spatial domain and may overlook informative cues in the frequency domain. In this paper, we propose SpectralPose, a frequency-enhanced pose estimation framework that integrates frequency-aware feature modulation with a heatmap refinement strategy to improve pose representation and keypoint localization. Specifically, intermediate feature maps are transformed into the frequency domain using the Fast Fourier Transform (FFT), where spectral gating and channel reweighting emphasize informative frequency components. The enhanced features are then projected back to the spatial domain via inverse FFT to enrich spatial representations. In addition, a lightweight refinement head learns residual corrections over predicted heatmaps to further improve localization precision. Experimental results on three benchmark datasets, including COCO, MPII, and OCHuman, demonstrate the effectiveness and robustness of the proposed approach. The source code is available at https://github.com/ted89229891/SpectralPose.

Keywords: Virtual and Augmented Reality Systems, Human Performance Modeling, Medical Informatics
Abstract: In the context of driving resumption for patients with unilateral spatial neglect (USN), evaluating cognitive function in dynamic environments is critical. To quantify USN-specific symptoms that conventional static assessments cannot fully capture, we developed a novel immersive virtual reality system to evaluate patients’ visual reactivity to moving objects (dynamic visual stimuli). Comparative experiments involving healthy individuals and patients with USN were conducted using this system. The preliminary results visualized cognitive characteristics specific to dynamic environments in the presented cases of USN patients, such as pronounced left–right asymmetry in reaction time and notable target omissions, which were not observed in the healthy group. The quantitative evaluation method using this system measures dynamic hazard-detection performance under demanding, driving-like conditions, which may inform more precise driving-aptitude assessment and rehabilitation strategies for USN patients.

Keywords: Medical Informatics, Machine Vision, Image Processing and Pattern Recognition
Abstract: In time-critical and resource-limited environments, clinicians have limited capacity to maintain accurate medical documentation while delivering critical care, making automated solutions essential for reliable clinical records and real-time data for managing triage, evacuation, and logistic resupply functions. Our recent programmatic experience coordinating automated field-care documentation research with multiple partners indicates that most existing approaches rely on object-centric pipelines requiring dense frame-level annotations, leading to long development cycles and limited robustness in realistic settings. Furthermore, these approaches remain poorly aligned with clinical documentation workflows, where annotations are inherently coarse and event-level, motivating learning under weak supervision. An object-agnostic weakly supervised multimodal framework is introduced that integrates vision–language–generated free-text descriptions with visual representations to infer applied medical procedures and treated body regions from egocentric video. Experimental results show superior performance over object-centric baselines with substantially reduced annotation effort, highlighting its scalability and suitability for real-world deployment.

Keywords: Medical Informatics, Human Performance Modeling, AI and Applications
Abstract: In this paper, we consider a smartphone application for automated physical function assessment based on the KOJI AWARENESS framework. The application is developed on Flutter with Google ML Kit BlazePose, so that all computation runs locally on the device. It covers 5 categories and 21 sub-items with bilateral scoring, using a two-phase evaluation protocol and per-landmark 1D Kalman filtering. A feasibility study with fourteen examinees confirms stable detection for 19 of the 21 sub-items at about 30~[FPS], while the shoulder mobility fails for all examinees due to known limits of BlazePose.

Keywords: Haptic Systems, Assistive Technology, Human Performance Modeling
Abstract: Research has shown that vibrotactile stimulation can enhance surgical task performance in minimally invasive surgery by increasing haptic sensitivity to compensate for limited haptic feedback. But its behavioral and neurophysiological effects may depend on stimulation location and duration. Behavioral performance was evaluated during a simulated laparoscopic tumor-discrimination task while recording synchronized EEG and force-torque data with a portable, low-cost setup. Where 24 participants completed four ordered sessions: pre-SR, SR, post-SR1, and post-SR2. Vibration was applied to functional locations, the instrument-operating arm, or to a non-functional, non-dominant arm, and results indicate significantly improved sensitivity proxy d-prime and accuracy in functional stimulation during SR, with d-prime remaining above baseline at post-SR2, whereas non-functional stimulation showed no significant improvement. This pattern suggests combined learning and time-location-dependent SR effects, although further group-by-session and EEG analyses are needed to draw concrete outcomes.

Keywords: Cooperative Work in Design, Game Theory and Multi-Agent Systems, Application of Artificial Intelligence
Abstract: Cooperation, a core aspect of human and animal behavior, is widely examined through game theory and models like the Prisoner's Dilemma. Understanding multi-agent cooperation reveals how individuals coordinate to achieve shared goals and inform practical applications. This work explores the potential of digital games as simulation environments for investigating social dilemmas and assessing cooperative behavior among groups. Focusing on paradigmatic scenarios — such as the n-person Prisoner's Dilemma and the Tragedy of the Commons — participants face choices between self-interest and collective benefit when managing shared resources. To address these challenges, we developed a customizable simulation in which players, acting as adaptive agents, learn from previous rounds and adjust their strategies accordingly. By systematically varying simulation parameters, our analysis reveals that cooperation often leads to more favorable outcomes than purely self-serving actions. These insights offer valuable practical applications for organizations simulating decision-making scenarios and predicting group behavior under diverse conditions.

Keywords: Game Theory and Multi-Agent Systems, Machine Learning, Distributed Intelligent Systems
Abstract: Limited work has examined the strategic behaviors of relational networked learning agents under social dilemmas, and has overlooked the intricate social dynamics of complex systems. We address the challenge with Socio-Relational Intrinsic Motivation (SRIM), which endows agents with diverse preferences over sub-graphical social structures in order to study the impact of agents' personal preferences over their sub-graphical relations on their strategic decision-making under sequential social dilemmas. Our results in the Harvest and Cleanup environments demonstrate that preferences over different subgraph structures (degree-, clique-, and critical connection-based) lead to distinct variations in agents' reward gathering and strategic behavior: individual aggressiveness in Harvest and individual contribution effort in Cleanup. Moreover, agents with different subgraphical structural positions consistently exhibit similar strategic behavioral shifts. Our proposed BCI metric captures structural variation within the population, and the relative ordering of BCI across social preferences is consistent in Harvest and Cleanup games for the same topology, suggesting the subgraphical structural impact is robust across environments. These results provide a new lens for examining agents' behavior in social dilemmas and insight for designing effective multi-agent ecosystems composed of heterogeneous social agents.

Keywords: Game Theory and Multi-Agent Systems, Networking and Decision-Making, Service Systems and Organizations
Abstract: Decentralized exchanges (DEXs) have become a cornerstone of decentralized finance (DeFi) due to their capital efficiency and market accessibility. However, liquidity fragmentation across multiple pools remains a significant challenge. In this paper, we propose a ``Smart Contract-as-a-Service'' (SCaaS) paradigm to address fragmentation by reusing existing DEX smart contracts as modular components for building DEX aggregators. We develop a time-dependent double auction mechanism for the DEX aggregator ecosystem, where liquidity providers (LPs) act as buyers, aggregator developers act as sellers, and the aggregator operator serves as the auctioneer. The auction incorporates time-varying valuations of liquidity provision, accounting for transaction fees and impermanent loss under dynamic pool conditions. A winner determination problem is formulated to optimize social welfare, and the mechanism is proven to satisfy individual rationality, budget balance, incentive compatibility, and computational efficiency. Simulation results demonstrate the effectiveness of the proposed mechanism in optimizing incentive structures and reducing liquidity fragmentation.

Keywords: Swarm Intelligence, Cooperative Systems and Control
Abstract: This paper presents a decentralized, force-based framework for fixed-size group formation in multi-agent systems. Agents rely only on local sensing to balance attraction and repulsion, allowing stable clusters of size K without centralized control or inter-agent communication during operation. A single communication step at initialization assigns unique identifiers that persist throughout the system. The method employs a logistic penalty function, parameterized by K and steepness, to regulate transitions between aggregation and repulsion, shaping group morphology and cohesion. A unified interaction model integrates free-agent motion, intra-group cohesion, and inter-group separation, while UUID-based identifiers provide robustness and scalability. Grouping quality is evaluated using the Silhouette Score Area Under the Curve (SS-AUC), a time-integrated clustering metric that captures both intra-group cohesion and inter-group separation across the full duration of group formation. The framework supports both homogeneous and heterogeneous populations through fully local, per-agent parameters, offering a flexible basis for decentralized task allocation, labor division, and spatial organization in swarm robotics and distributed autonomous systems.

Keywords: Networking and Decision-Making, Game Theory and Multi-Agent Systems
Abstract: In extreme emergency scenarios with paralyzed ground infrastructure and dynamic interference, low-altitude unmanned aerial vehicles (UAVs) struggle to maintain reliable service due to limited battery capacities. To address this, we propose a High Altitude Platform (HAP)-UAV collaborative air-ground network. This architecture leverages HAPs' wide-area coverage and line-of-sight transmission as a robust anti-jamming baseline, while introducing Wireless Power Transfer (WPT) to replenish UAV energy and overcome hovering duration bottlenecks. To tackle the severe intra- and inter-cell interference under Non-Orthogonal Multiple Access (NOMA) coupled with multi-dimensional resource constraints, we formulate the joint optimization problem as an Anti-jamming Task Offloading and Deployment (ATOD) game. By constructing a four-dimensional generalized potential function encompassing user costs, computation, UAV energy, and interference penalties, we prove the existence of a Nash Equilibrium (NE) and design the Distributed Anti-jamming Resource Equilibrium (DARE) algorithm. Simulation experiments verify the effectiveness of the DARE algorithm, showing promising performance in achieving fast convergence and minimizing system cost.

Keywords: Multimedia Systems, Media Computing, Human Perception in Multimedia
Abstract: 3D Gaussian Splatting (3DGS) has recently emerged as an efficient and expressive representation for view synthesis and object modeling. However, its application to object retrieval remains challenging because each object is represented as a variable-sized set of Gaussians and currently available datasets are often limited in scale. This paper proposes a denoising autoencoder (DAE)-based framework for augmentation and canonicalization for 3DGS object retrieval. The proposed framework reconstructs Gaussian attributes from partially corrupted inputs and produces identity-preserving local variants around the original data manifold, rather than synthesizing entirely novel objects. We investigate augmentation strategies with different appearance perturbation strengths and examine their role in retrieval-oriented representation learning under limited-data conditions. Experiments on 87 synthetic 3DGS objects show that the proposed pipeline can systematically generate identity-preserving augmented variants and construct original–augmentation pairs for retrieval-oriented learning. The current evaluation should be interpreted as a pilot feasibility study that demonstrates pair consistency and optimization stability, rather than as evidence of full retrieval generalization. These results suggest that DAE-based augmentation is a promising component for future 3DGS retrieval systems under limited-data conditions.

Keywords: Generative AI, Communications, Application of Artificial Intelligence
Abstract: Classical packet-erasure recovery targets exact reconstruction, but when full recovery begins to fail, structured payloads may still support useful partial recovery. Natural language provides a strong test case because surviving context can help infer missing content. We introduce GenFEC, a constrained generative recovery method that transmits lightweight deterministic token-bucket labels and uses receiver-side language-model decoding to resolve ambiguity at erased positions while freezing surviving tokens. We evaluate 60 fixed-length WikiText-2 messages over IID, burst, and Gilbert–Elliott erasure channels. Under matched measured overhead and a shared source/channel harness, controlled classical references remain better on exact-message recovery at 4, 8, and 16 metadata bits per source token. GenFEC nevertheless exhibits a graceful-degradation regime: its best overall exact erased-token recovery configurations improve that metric by 13.2 and 9.2 percentage points at 8 and 16 bits, with message-level 95% intervals entirely positive, while trailing at 4 bits. A frozen frequency-aware tiered_hash map further improves recovery at unchanged transmitted bit width by reducing expected candidate ambiguity for common in-domain tokens. Sentence-embedding similarity provides supporting within-family evidence, but current GenFEC inference takes seconds rather than sub-milliseconds. These results position GenFEC as a complementary compute-for-redundancy layer for delay-tolerant structured text, not a replacement for conventional coding.

Keywords: Human-Machine Interface, Human Factors, Haptic Systems
Abstract: Robotic grasping and manipulation tasks in the real-world often require human monitoring, as autonomous systems struggle to handle the complexities of the environment. Human monitors utilize live camera feeds, but these can often be degraded due to environmental factors. Tactile sensing offers a promising complement to vision; however, the extent to which visualizations of tactile information can support human monitors under visually-constrained environments requires further investigation. To better understand the utility visual displays of force feedback provide in these environments, we employed a prototype tactile sensor alongside a camera on a puppeted robot gripper to generate and present tactile grasp data and a degraded camera feed to users. We evaluate how different tactile visualizations impact performance on a grasp monitoring task compared to when only degraded (low quality) camera feed is available. Eye tracking was used to assess attention to camera versus tactile information. Results suggest that tactile visualizations can provide information to combat Type 1 errors (stopping what would have been a successful grasp), but may also distract operators when the camera feed alone is sufficient for judgment. These findings highlight the need to carefully consider when and how tactile information is presented, as it has the potential to both improve and impair human supervision of grasps within visually-constrained environments.

Keywords: Brain-Computer Interfaces, Deep Learning, Medical Informatics
Abstract: Mild traumatic brain injury (mTBI) is a prevalent condition that remains difficult to diagnose in its early stages. Oculomotor dysfunction is a well-established marker of mTBI, motivating the development of portable tools that capture both eye-movement behavior and underlying neurophysiology. In this work, we present an initial framework that integrates electroencephalogram (EEG) with augmented-reality (AR)-based Vestibular/Ocular Motor Screening (VOMS) tasks to estimate subject-specific ocular response times. Pre-processed EEG signals, obtained through band-pass filtering and average referencing, are analyzed using a Redundant Discrete Wavelet Transform (RDWT)–driven deep neural framework. The RDWT coefficients are subjected to trainable zero-phase convolutional filtering and reconstructed into the time domain via inverse RDWT, followed by channel-wise temporal and spatial filtering using 2D convolution layers and convolutional–LSTM-based decoding. An ablation study demonstrates that wavelet-domain filtering serves as an effective denoising strategy, improving prediction performance. Sliding-window predictions were validated using Pearson correlation (> 0.5), and Dynamic Time Warping (DTW) was subsequently used to estimate ocular response times. DTW-derived metrics revealed VOM-task-dependent differences in ocular response time between participants, characterized using Mann–Whitney U tests. Cross-correlation analysis further revealed task-dependent temporal behaviors: pursuit tasks exhibited reactive tracking, whereas saccades showed anticipatory responses. Overall, the results highlight pursuit tasks as particularly informative for distinguishing timing differences and demonstrate the potential of RDWT-based EEG features combined with DTW metrics for multimodal mTBI assessment.

Keywords: Robotic Systems, Mechatronics
Abstract: This paper presents VacuBot, a 10-DOF reconfigurable wheel-legged robot designed to bridge the gap between high-speed wheeled transit and versatile legged manipulation. Rather than relying on heavy auxiliary robotic arms, VacuBot repurposes its primary limb kinematics and distal vacuum-adhesion modules to perform task-space manipulation. We establish a unified dynamic model using the Lagrange-Euler formalism to capture the floating-base constraints during morphological transitions. To handle unmodeled contact forces, we deploy a hierarchical control architecture featuring a task-space impedance controller. Experimental results demonstrate that the proposed framework effectively dampens center-of-mass perturbations during morphological switching and maintains stable locomotion while executing 1.5 kg payload manipulation.

Keywords: Generative AI, Machine Learning, Decision Support Systems
Abstract: Predictive process monitoring traditionally relies on models trained on large volumes of historical event logs, limiting deployment in low-data or rapidly evolving workflows. We study if large language models (LLMs) can serve as training-free predictors for next-event prediction while respecting process control-flow constraints. We introduce Grammar Violation Rate (GVR), a structural compliance metric that quantifies whether predicted transitions conform to process-mined execution grammars. Evaluating three LLMs — Claude Haiku 4.5, Claude Sonnet 4.5, and GPT-4.1-nano — on two real-world event logs, we show that zero-shot prompting produces substantial structural violations despite moderate accuracy. However, retrieval-augmented few-shot prompting reduces violations to near zero using as few as five examples while substantially improving prediction accuracy. On a complex clinical pathway dataset, retrieval-augmented LLMs surpass trained Markov and nearest-neighbor baselines and approach the performance of a trained LSTM without any task-specific training. These results demonstrate that structurally relevant in-context examples enable LLMs to recover process-specific transition constraints, supporting training-free predictive monitoring in data-limited environments.

Keywords: Brain-Computer Interfaces, Human-Machine Interaction, Deep Learning
Abstract: The decoding of linguistic information from electroencephalography (EEG) signals remains an extremely challenging problem in brain-computer interface (BCI) research. In particular, sentence-level decoding from EEG is difficult due to the low signal-to-noise ratio (SNR) of these recordings. Previous studies tackling this problem have typically failed to surpass random baseline performance unless teacher forcing is used during the inference phase. In this work, we propose a retrieval-augmented generation (RAG)-based sentence-level EEG-to-text decoding pipeline that combines an EEG encoder aligned with semantic sentence embeddings, a vector retrieval stage, and a large language model (LLM) to refine retrieved sentences into coherent output. Experiments are conducted on the Zurich Cognitive Language Processing Corpus (ZuCo) dataset, which contains single-trial EEG recordings collected during silent reading. To evaluate whether the system extracts meaningful information from these EEG signals, the results are compared with a random baseline. In nine subjects, the proposed pipeline outperforms the random baseline, achieving a mean cosine similarity of 0.181 pm 0.022 compared to 0.139 pm 0.029 for the baseline, corresponding to a relative improvement of 30.45%. Statistical analysis further confirms that this improvement is significant, following a strict evaluation workflow where inference is performed without access to ground-truth labels.

Keywords: Active BMIs, Brain-Computer Interfaces, BMI Emerging Applications
Abstract: Brain–machine interfaces (BMIs) can assist individuals with limited mobility, such as stroke survivors or amputees. One of the key challenges in developing BMIs is expanding their usability and control, which can be achieved by accurately decoding multiple kinematic and kinetic parameters. To address this, we propose three regression models—partial least squares regressor, multilayer perceptron, and attention-based regressor—to decode multiple movement parameters from EEG signals. We evaluated these models using the WAY-EEG-GAL dataset, focusing on their performance under subject-specific and subject-independent conditions with two strategies: a single model for all parameters and a baseline with separate models for each parameter. Among all regressors, the attention-based regressor achieved the best performance, with an R^2 of 0.8 and a latency of 29.2 milliseconds, demonstrating significant improvement in simultaneous multi-parameter decoding, although its performance dropped for single-parameter decoding. The multi-layered perceptron showed more consistent but lower accuracy across both decoding types (R^2 = 0.49). These findings highlight the potential of attention-based models for real-time multi-command BMI systems and contribute to the development of more intuitive control devices.

Keywords: Active BMIs, Brain-Computer Interfaces, BMI Emerging Applications
Abstract: Brain--machine interfaces provide a link between neural activity and external devices, enabling restoration of motor function and advancing human--machine interaction using non-invasive electroencephalography (EEG). However, continuous grasp force decoding remains challenging due to complex temporal dynamics, high inter-subject variability, and limited generalisation of existing approaches. To address this, we propose a hybrid EEG decoding framework that jointly models continuous and tokenised representations, enabling capture of both fine-grained neural structure and long-range temporal dependencies. The proposed approach integrates convolutional--recurrent representation learning, quantisation-based tokenisation, and transformer-based temporal modelling within a unified fusion-based regression architecture. Experimental evaluation on the WAY-EEG-GAL dataset under strict leave-one-subject-out conditions achieves R^2 = 0.817 in offline settings and R^2 = 0.793 in simulated real-time evaluation, with latency suitable for real-time deployment. These results demonstrate strong cross-subject generalisation and highlight the practicality of hybrid continuous--tokenised representations for real-time EEG-based force decoding in assistive robotics, neuro-rehabilitation, and human--machine interaction.

Keywords: Brain-Computer Interfaces, Brain-based Information Communications, Representation Learning
Abstract: Generalizing across subjects remains challenging in invasive neural recordings because electrode configurations, anatomical structures, and neural signal patterns vary substantially across individuals. To investigate such inter-subject variability, we propose a cross-subject semantic decoding framework that aligns neural responses to speech perception from multiple subjects into a shared latent space and learns a mapping from the aligned neural representations to contextual embeddings. More specifically, using electrocorticography data collected during natural language comprehension, we estimate the shared space using the shared response model and train a decoder to predict contextual semantic embeddings from projected neural responses. For a held-out subject, we estimate a subject-specific projection into the predefined shared space, and directly apply the pretrained decoder without any retraining. Experimental results demonstrate that the proposed framework consistently outperforms baseline methods across evaluation settings and exhibits a reduced performance drop from source subject to held-out subject testing, indicating improved cross-subject generalization. These results suggest that aligning neural activity into a shared latent space, while decoding in a semantic embedding space, provides an effective strategy for improving cross-subject generalization by reducing subject-specific differences in neural responses while effectively capturing shared stimulus-related representations.

Keywords: Brain-Computer Interfaces, Machine Learning, Information Systems for Design
Abstract: Performance in brain–computer interfaces (BCIs) is commonly summarized using metrics that combine speed and accuracy, masking how precision evolves as evidence accumulates during inference. This limitation is critical under low-evidence conditions, where early-stopping strategies in P300-based BCIs reduce response time by imposing an intrinsic speed–accuracy trade-off.

We characterize this dependency using Gain–Conservation trade-off curves, describing accuracy across evidence constraints relative to a baseline. From these curves, we derive the partial area under the curve to summarize regime-specific performance and the evidence sensitivity to quantify sensitivity to evidence variations, explicitly separating low- and high-evidence behavior. We further introduce a parametric formulation in which an evidence scaling exponent λ captures low-evidence efficiency as performance improves with accumulating evidence.

Applied to multiple classifiers and paradigms, the framework reveals regime-dependent differences in evidence utilization and accuracy stability, with larger differences emerging under low-evidence conditions. The exponent λ provides a summary of evidence efficiency, where low values indicate effective evidence use and high values greater reliance on repeated trials.

Overall, this framework provides an evidence-aware methodology for evaluating classifier stability and paradigm design in P300-based BCIs under limited-evidence conditions.

Keywords: Brain-based Information Communications, Brain-Computer Interfaces, Information Visualization
Abstract: Decoding visual stimuli from brain signals remains challenging due to the presence of multi-level latent representations and complex functional interactions across brain regions. Existing approaches often rely on sequence similarity between regions, which provides limited capacity for modeling medium- to long-range functional associations. Graph-based methods offer a promising alternative by capturing topological relationships, but many existing graph-based approaches overlook temporal dependencies between samples and require cumbersome preprocessing, limiting their generalizability.In this paper, we propose a novel visual decoder-guided framework that explicitly models spatiotemporal functional connectivity. Treating fMRI signals as time-series data, we employ patch-based sequence extraction to compute correlation coefficients and construct a global topological graph. An encoder is then used to extract spatiotemporal graph features, which are fused with fine-grained sparse mask features to guide image reconstruction. By integrating graph-theoretic topological representations with feature-level fusion, the proposed framework effectively leverages functional connectivity for visual decoding.Experimental results on the GOD and BOLD5000 datasets demonstrate that our method consistently outperforms existing baselines, achieving superior reconstruction quality and improved image clarity.

Keywords: Systems Safety and Security, Distributed Intelligent Systems
Abstract: The rapid deployment of multi-agent systems (MAS) in financial decision-making introduces critical systems safety challenges. Focusing on corporate credit rating, this work empirically evaluates a hierarchical MAS against four runtime perturbations: data poisoning, memory poisoning, and both observational and direct prompt injections. By tracking how localized deviations cascade through the agent network, we reveal three key findings: (1) Asymmetric Vulnerability Across Rating Tiers: higher-rated entities exhibit disproportionately high sensitivity to adversarial injections compared to lower-rated counterparts; (2) Divergent Propagation Topologies: localized errors trigger either direct proportional propagation or latent cascading and delayed amplification, fundamentally shaped by the perturbed agent's functional role during consensus aggregation; (3) Empirical Resilience: a modular recovery framework successfully intercepts anomalous propagation, restoring up to 93.2% of system-level ratings.

Keywords: Evolutionary Computation, Metaheuristic Algorithms, Computational Intelligence
Abstract: 动态多目标优化问题（DMOPs）包括 在资源有限的实时环境中广泛遇到 比如物联网（IoT）边缘计算等场景。 现有的基于预测的动态多目标 进化算法常常面临相当大的困难 响应开销，因为复杂的预测 机制通常依赖于复杂的自适应建模或 密集的历史国家管理。为了解决这个问题 本论文提出了一个动态多目标方案 基于双分支预测和 轻量级聚类突变，称为DBP-LCM。 具体来说，DBP-LCM结合了全局质心漂移和 局部边界引导信息以生成轻量级 环境变化后预测的种群。前往 进一步缓解预测引起的局部拥挤，a 引入了基于聚类的多样性修复策略， 其中对冗余应用自适应柯西变异 通过密度感知聚类识别的区域。在这方面 该方法改善了种群重新分布 同时保持低响应开销

Keywords: Autonomous Vehicle, Deep Learning, Intelligent Transportation Systems
Abstract: To address the problem of insufficient modeling of the coupling relationship between intent and trajectory in lane change behavior for autonomous driving, this paper proposes an Intent-Guided Attentive Social Modulation Framework (IGASM). The framework extracts intent vectors through social attention and introduces Feature-wise Linear Modulation (FiLM) and a soft gating mechanism during the trajectory decoding stage to achieve joint modeling of intent recognition and trajectory prediction. A lane change dataset is constructed based on the NGSIM dataset for experiments. Under both LSTM and Transformer frameworks, the proposed method outperforms the baseline in metrics such as accuracy for the intent recognition task. For the trajectory prediction task, metrics including MSE are reduced by 7% to 37% compared to the baseline. Comparative experiments and ablation studies validate the effectiveness of the proposed method.

Keywords: Systems Safety and Security, AI and Applications, Expert and Knowledge-Based Systems
Abstract: With the rapid development of Large Language Model (LLM) App Stores, the surge in customized LLM apps has become a significant trend. However, this expansion has also brought severe security challenges. Although existing studies have analyzed the security risks in LLM App Stores, their focus is often limited to the LLM apps themselves, failing to cover the store’s publication workflow. We investigated the ecosystem security of LLM App Stores from the perspective of the LLM app publication workflow. We developed an automated and store-independent framework that can reproduce malicious publisher behaviors, batch create malicious LLM apps, bypass store audit mechanisms, and record and evaluate behavior after publication. We selected mainstream app stores including Coze, Baidu Wenxin, and Poe, built a multimodal malicious knowledge base containing high-risk text and visual content, and deployed three types of LLM apps (text-based, image-based, and plugin-based) for testing. The evaluation was carried out from multiple dimensions such as output harmfulness, attack success rate, semantic fidelity, and text-image consistency, and revealed systemic vulnerabilities in the knowledge injection layer of the LLM application ecosystem: 1) Malicious or harmful knowledge bypasses filtering mechanisms and is re-outputted as is or with high semantic fidelity; 2) Enabling retrieval tool plugins significantly amplifies the risk potential of harmful content output; 3) In multiple malicious domains, Attack Success Rate (ASR) initiated through image input is close to 100%, and the prompt is highly consistent with the image. To facilitate reproducibility, we have made our implementation publicly available at the project repository.

Keywords: Generative AI, System Architecture, Expert and Knowledge-Based Systems
Abstract: Long-context agents often fail when answer-critical evidence is sparse, weakly localized, or unavailable in the active prompt. Memory-augmented agents address this problem with retrieval, caching, hierarchical storage, and adaptive context assembly, but these mechanisms are often evaluated together, making it difficult to isolate the contribution of persistent retrieval. We evaluate memory-layer composition under a fixed execution stack in which the base model, decoding, prompt scaffold, parser, retrieval scorer, and budget policy remain unchanged. Five variants are compared under two support conditions: evidence-sufficient inputs keep answer-bearing support in the active prompt, whereas evidence-scarce inputs relocate the same support to retrievable memory. Across LoCoMo, LongBench~v2, and LongMemEval, current-turn organization (B1) achieves the highest score among the tested variants in all evidence-sufficient runs. Under the evidence-scarce protocol, L0+L2 retrieval (V1) achieves the highest score on all three datasets and improves over B1 by +0.0628, +0.1233, and +0.1220 absolute score points, respectively, with an unweighted mean gain of +0.103. These results indicate that, in the evaluated setting, persistent L2 retrieval is beneficial when answer-bearing support is removed from the active prompt, while adding L1/L3 on top of L2 does not improve the best observed scarce-regime result.

Keywords: Generative AI, Machine Learning
Abstract: The use of large language models (LLMs) for article generation has become common practice. However, limited research has been conducted on writing expository texts, and existing methods often require substantial human involvement, leading to low efficiency and weak structural coherence. By emulating human article editing workflows, we propose MUSE, a Multi-agent Unified Structured Expository writing Framework. MUSE aligns the LLM-based multi-agent long-form text generation process with the human writer's writing process, automatically generating and editing explanatory texts through three steps: planning, generation, and proofreading. Additionally, it incorporates reader agents and editor agents to iteratively refine the generated content, while retaining the experience gained during the revision process for use in subsequent writing. We conduct comprehensive experiments using both automated and human evaluation of the generated articles. The articles produced by MUSE exhibit clear advantages in writing efficiency and overall quality, and achieve significant improvements in factuality, accuracy, and coherence compared with other text generation frameworks.

Keywords: Distributed Intelligent Systems, Machine Learning, Deep Learning
Abstract: Personalized federated learning (PFL) aims to tackle client heterogeneity by learning customized models for each client rather than enforcing a single global solution. However, existing approaches typically focus on either representation alignment or classifier collaboration in isolation, limiting their effectiveness under severe non-IID conditions where feature and decision heterogeneity are tightly coupled. In this paper, we propose FedSAA, a unified personalized federated learning framework that explicitly decouples representation-level alignment from decision-level collaboration. Specifically, FedSAA introduces semantic anchor alignment to stabilize local feature learning across heterogeneous clients. In addition, a similarity-aware head aggregation mechanism enables each client to selectively collaborate with its most relevant peers. This design mitigates unreliable similarity estimation caused by feature drift and reduces negative transfer without enforcing global classifier consensus. Extensive experiments on CIFAR-10 and CIFAR-100 under various non-IID settings demonstrate that FedSAA consistently outperforms state-of-the-art baselines, including FedAvg, FedPAC, and FedAS. Notably, the proposed method achieves significant performance gains on the fine-grained CIFAR-100 dataset, highlighting its robustness in scenarios with severe heterogeneity. Further ablation studies validate the distinct contributions of our decoupled design and show that selectively collaborating with a small number of similar clients yields a better trade-off between personalization and knowledge sharing compared to global aggregation.

Keywords: Game Theory and Multi-Agent Systems, Agent-Based Modeling, Generative AI
Abstract: Large Language Models (LLMs) have been in wide use in various fields in recent years. Despite their strong performance on many individual tasks, however, no single LLM is universally optimal across diverse tasks and domains. Practical systems often rely on manually orchestrating multiple LLMs for different types of tasks. Moreover, since these models operate independently and typically cannot share context, such multi-LLM setups often suffer from inefficient and fragmented reasoning. To bridge this gap, we present a unified multi-LLM framework named Automatic Task-aware Context-preserving Routing (ATCR), which automatically routes tasks to the most suitable model and maintains a shared global context by summarizing prior interactions into a compact, structured context prompt that is prepended to subsequent calls, enabling more efficient and context-consistent task execution. In addition, we employ a lightweight evolutionary loop guided by programmatic scoring and LLM-based evaluation to further refine the final output. Experiments on a custom cross-modal benchmark that we construct from real usage scenarios show that our framework improves performance by 3.3 points over a no-agent baseline in Easy mode and slightly reduces average cost according to cost logs, while Hard mode refinement yields larger gains at higher cost. Moreover, our approach consistently outperforms random routing baselines across a broad set of benchmarks and evaluation metrics, including MMLU-Pro and MMMU-Pro.

Keywords: Machine Vision, Image Processing and Pattern Recognition, Deep Learning
Abstract: Self-supervised monocular depth estimation has emerged as a promising alternative to supervised learning by removing the need for dense ground-truth depth. However, existing lightweight methods usually treat decoding as a uniform feature restoration process, while the dominant errors in self-supervised depth estimation emerge progressively across decoding stages. Early upsampling tends to dilute semantically reliable responses, intermediate cross-scale fusion is vulnerable to background interference, and late high-resolution reconstruction often lacks explicit boundary-sensitive guidance. To address this issue, we propose a lightweight self-supervised monocular depth estimation model with a multi-stage collab orative guided decoder. Instead of stacking several generic enhancement modules, the proposed decoder is organized as a stage-matched recovery pipeline: CRU preserves semanti cally rich responses during early upsampling, GCGB performs structure-aware cross-scale fusion, PCSAB conducts selective contextual refinement, and FEGM injects explicit foreground edge guidance into high-resolution reconstruction. This design enables progressive recovery of decoder-side information loss under a compact computational budget. Extensive experiments demonstrate that the proposed method achieves competitive performance while maintaining an excellent balance between estimation accuracy and computational efficiency.

Keywords: Machine Vision, Image Processing and Pattern Recognition, Transfer Learning
Abstract: Referring expression comprehension (REC) aims to localize a target region in an image given a natural language description. Adapting open-vocabulary detectors to REC through serial parameter-efficient fine-tuning (PEFT) risks disrupting the pretrained feature-head coupling in DETR-style architectures. Moreover, these detectors exhibit bag-of-words behavior, focusing on dominant nouns while neglecting fine-grained attributes and spatial relations. To address these challenges, we propose GRAV, a two-stage framework comprising a Decoupled Grounding Adapter (DGA) and a Multimodal Large Language Model (MLLM)-Assisted Verification (MAV) module. DGA is designed as a parallel side branch that preserves the backbone's feature integrity through zero-initialization and gradient isolation. Within this branch, multimodal features are decoupled into orthogonal subspaces separating entity semantics from relational cues. These components are then aligned through distribution matching and optimal transport, respectively, effectively mitigating the bag-of-words bias. To further resolve ambiguities beyond feature-level alignment, MAV employs a frozen MLLM with an early-exit mechanism to verify top candidates without additional training. With only 0.44M trainable parameters (0.19% of the backbone), GRAV consistently improves Grounding DINO across RefCOCO/+/g benchmarks, achieving accuracy competitive with models that are over 16,000× larger.

Keywords: Generative AI, Application of Artificial Intelligence, Image Processing and Pattern Recognition
Abstract: Recent retrieval-augmented image captioning methods have made progress in improving image understanding. They incorporate external knowledge to compensate for the limitations of single-image models in comprehending complex scenes. However, current approaches face challenges in relation modeling: (1) the representation of semantic prompts is too coarse-grained to capture fine-grained objects and their relationships; (2) these methods lack explicit modeling of image objects and their semantic relationships, which results in limited expressiveness at the level of object relations. To address these limitations, we propose RACap, a relation-aware retrieval-augmented model for image captioning. Our model not only mines structured relation semantics from retrieval captions, but also identifies heterogeneous objects from the image. Then RACap effectively retrieves structured relation features that contain heterogeneous visual information and integrates them into the image representation to enhance the semantic consistency and relational expressiveness. Experimental results show that RACap, with only 10.8M trainable parameters, achieves superior performance compared to previous lightweight captioning models on datasets such as COCO, Flickr30k, and NoCaps.

Keywords: Metaheuristic Algorithms, Logistic and Manufacturing Automation, Application of Artificial Intelligence
Abstract: Improving task coverage while satisfying individual travel-budget constraints for unmanned aerial vehicles (UAVs) remains a challenging problem in multi-UAV mission planning. This paper addresses a range-constrained multi-UAV routing problem that aims to maximize the total value of served tasks subject to individual UAV distance limits. We propose a multi-agent collaborative adaptive large neighborhood search (MA-ALNS) framework that integrates heterogeneous parallel search, adaptive meta-synchronization, heuristic phase switching, and priority-based reflow of unserved tasks. The heterogeneous agents explore the solution space with complementary search behaviors, while adaptive meta-synchronization selectively transfers high-quality route structures among agents. Priority-based reflow identifies tasks that remain unserved and attempts to reinsert them into feasible routes, thereby improving task coverage, particularly under tight range constraints. Experiments on large-scale random instances under standard and tight constraint settings show that MA-ALNS consistently achieves higher mean objective values than the genetic algorithm (GA), particle swarm optimization (PSO), and all evaluated ablation variants across every combination of problem scale and constraint setting. The performance gains are more substantial under tight constraints. The ablation results further show that multi-agent parallel search establishes the performance foundation, while priority-based reflow provides the largest improvement in task coverage.

Keywords: Machine Learning, Deep Learning, Representation Learning
Abstract: Traditional sequential recommendation methods rely on a single embedding to represent users, limiting their ability to model diverse preferences. Although multi-interest models address this issue by learning multiple interest vectors, they treat these vectors as mutually independent, overlooking the relational structures among them, particularly the synergistic alignment and mutual suppression between interests. We propose IRGRec, a sequential recommendation framework that explicitly models inter-interest dependencies via dynamically constructed relation graphs. A hybrid collaborative-semantic similarity constructs a dual-channel graph that simultaneously reinforces compatible interests and suppresses conflicting ones, enabling the model to leverage the structural topology of the interest space during candidate generation. Extensive experiments on three public benchmarks demonstrate that IRGRec consistently outperforms strong baselines, validating the effectiveness and superiority of explicitly modeling inter-interest dependencies.

Keywords: Distributed Intelligent Systems, Machine Learning, Representation Learning
Abstract: ，个性化联邦学习面临关键稳定性 数据异质性下的挑战，即交替出现 局部-全局优化诱导易变特征 表征。我们观察到这一被忽视的现象 作为表示不一致的累积 几何色散引起的波动性与 时间漂移，这会降低系统收敛性， 个性化质量。 为解决这个问题，我们提出了FedHypSA，一个 闭环几何-时间稳定框架。 超球形批次监督对局部有限制 特征映射到超球形流形上，确保 类内紧致性与类间可分离性。 自动感知级对准机制 通过闭环反馈调节比准强度 来自一种新的时空波动率指标。 CIFAR-10、CIFAR-100 上的广泛实验， Tiny-ImageNet 和 Office10 展示了更优的优势 性能显著提升了稳定性 最先进的方法，显著地证明了 在严重数据中减少表征振荡 异质性设置，验证稳健稳定 分布式异构Ŀ

Keywords: Image Processing and Pattern Recognition, Medical Informatics, Deep Learning
Abstract: Colorectal cancer is one of the most common malignant tumors of the digestive tract, and early detection is crucial for improving patient survival rates. Lesion detection based on CT images plays a vital role in early screening, disease assessment, preoperative planning, and therapeutic monitoring. However, due to the irregular morphology, blurred boundaries, and low contrast of lesions, both manual interpretation and existing object detection algorithms face significant challenges in colorectal lesion identification, manifested by reduced detection accuracy and a high rate of missed detections for small lesions. To address these issues, this study proposes DSAH-YOLO integrated with an Adaptive Super-Resolution Enhancement (ASRE) for accurate lesion detection in colorectal CT images. The ASRE enhances the visual quality of abdominal CT patches, effectively mitigating issues caused by low-resolution imaging such as blurred lesion boundaries, loss of structural details, and insufficient feature extraction. Furthermore, a combination of Dynamic Multi-View Feature Fusion (DMVFF) and Multi-Scale Local Feature Enhancement Attention (MS-LFEA) is designed to improve the perception of lesion edges and deep structural features. In addition, a Dual-Branch Feature Guided Detection (DBFGD) Mechanism is reconstructed to optimize the detection of small-area and low-contrast lesions. Experimental results demonstrate that the proposed DSAH-YOLO model achieves consistent performance improvements over the baseline model. Specifically, the proposed method improves precision by 9.1%, sensitivity by 5.4%, and mean mAP@0.5 by 5.7% on the CC-CT-V dataset, maintaining superior detection performance compared to other end-to-end models. This approach provides real-time diagnostic assistance for clinicians in early colorectal cancer screening and offers a technical foundation for quantitative lesion analysis.

Keywords: Modeling of Autonomous Systems, Energy Systems, Intelligent Power Grid
Abstract: There exists a complex nonlinear coupling relationship among electrical, cooling, and heating loads in integrated energy systems. Accurate joint forecasting of multi-energy loads is an important prerequisite for optimal dispatching and economic operation of the system. This paper proposes a short-term multi-energy load joint forecasting model, CLA-FDKformer, based on an improved iTransformer architecture, aiming to enhance the joint forecasting accuracy of electrical, cooling, and heating loads. First, the iTransformer inverted embedding strategy is introduced to address the problem that multivariate mixed modeling in traditional methods destroys the physical independence of individual loads. Second, a Frequency-Domain Decomposition KAN module (FD-KAN) is proposed to enhance the model's ability to capture nonlinear temporal features of multi-energy loads through frequency-domain decoupling and Kolmogorov-Arnold Networks with learnable B-spline activation functions. Finally, a Cross-Load Attention module (CLA) is designed to explicitly model the physical coupling relationships among multi-energy loads by constructing directed cross-attention paths and a gated adaptive fusion mechanism. Experimental results demonstrate that CLA-FDKformer outperforms multiple baseline methods across all three load forecasting tasks, providing reliable forecasting support for the optimal dispatching and safe economic operation of integrated energy systems.

Keywords: Communications, Deep Learning, Complex Network
Abstract: Cellular traffic prediction is important for proactive network management and resource allocation. Real-world cellular traffic exhibits pronounced spatio-temporal heterogeneity and non-stationarity, where inter-cell dependencies vary with node--time context. Existing methods either ignore contextual information or use it only as an auxiliary feature fused into traffic representations. They fail to explicitly use contextual information to model context-dependent spatial interactions. We propose CASTLE, a context-aware spatio-temporal attention network with a Contextual Spatial Block (CSB) that treats node--time context as an explicit control signal. Specifically, CSB performs context-conditioned relation learning to determine which inter-cell dependencies should be activated under different contexts, and uses context-conditioned update gating to control how much neighborhood information is written into node representations. Extensive experiments on the real-world Telecom Italia datasets show that CASTLE consistently outperforms strong baselines.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Rice panicle detection is critical for crop growth monitoring and yield estimation, yet the small size of rice panicles and complex field backgrounds hinder efficient real-time detection in unmanned aerial vehicle (UAV)-based agricultural scenarios. Conventional detection models suffer from high computational consumption and poor real-time performance, limiting their field application value. To solve this problem, this paper proposes a lightweight high-speed SBC-RTDETR detection framework based on RT-DETR. The embedded Space-to-Depth Convolution (SPDConv) module compresses spatial information and enhances fine-grained feature extraction while reducing computational complexity. The Bidirectional Feature Pyramid Network (BiFPN) structure optimizes multi-scale feature fusion, and the Content-Aware ReAssembly of Features (CARAFE) module preserves spatial details during upsampling with negligible computational overhead. The model is trained on a high-quality UAV rice panicle dataset collected from Yongchuan, Chongqing. Experimental results show that compared with the baseline model, SBC-RTDETR reduces model parameters and computation by 37%, and boosts the inference speed from 31.3 FPS to 50.1 FPS (a 66.4% improvement) under the same hardware conditions. Meanwhile, the mAP@0.5 and recall are slightly improved by 0.8% and 1.0%, respectively, realizing high-speed inference without accuracy loss. The proposed SBC-RTDETR achieves efficient small-target detection in complex field environments and is well-suited for real-time deployment in UAV-based smart agriculture systems.

Keywords: Image Processing and Pattern Recognition, Representation Learning, Machine Learning
Abstract: Open-vocabulary object detection (OVOD) aims to detect objects from novel categories beyond the predefined training set, enabling more flexible visual understanding in real-world scenarios. Recent approaches typically rely on distilling knowledge from vision-language models (VLMs) to align visual and textual representations. However, their performance remains limited when handling challenging instances, such as small, blurry, or visually ambiguous objects. In this paper, we propose CGVD (Caption-Guided Classification and Variational Query Mapping for Open-Vocabulary Object Detection), a novel framework that enhances OVOD by incorporating caption-level semantics and probabilistic visual-to-language mapping. First, we introduce a caption-guided classification refinement module, which leverages global contextual information from image captions to refine instance-level predictions. Second, we design a VAE-based visual-to-language projector that models uncertainty in visual queries and produces more robust textual representations for cross-modal alignment. We evaluate our method on two widely used benchmarks, OV-COCO and OV-LVIS. Experimental results show consistent improvements over a strong baseline, with more pronounced gains on challenging instances such as small objects. These results suggest that combining caption-level context with probabilistic feature modeling can improve robustness and generalization in open-vocabulary object detection.

Keywords: Deep Learning, Human-Machine Interaction, Human-Computer Interaction
Abstract: Graph Convolutional Networks (GCNs) often struggle to balance recognition accuracy with computational efficiency. To bridge this gap, we propose the Dynamic Multi-Scale Selective Graph Convolutional Network (DMS-GCN), a lightweight yet powerful framework for skeleton-based action recognition. DMS-GCN features a Dynamic Dual-scale Spatial Modeling (S-GCN) module that adaptively fuses multi-scale semantic information via anatomical hypergraph priors and a scale-selection gate. Complementarily, an Adaptive Multi-scale Temporal Modeling (T-GCN) module employs structured sampling to decouple the observation window from the parameter count, enabling flexible adaptation to diverse action rhythms while preserving temporal causality. Furthermore, a lightweight spatio-temporal channel attention mechanism is integrated to refine global contextual features. Experimental results on NTU RGB+D and NTU RGB+D 120 datasets demonstrate that DMS-GCN achieves state-of-the-art performance with significantly reduced computational overhead, offering a superior trade-off for resource-constrained deployments.

Keywords: Cognitive Computing, Medical Informatics, Computational Life Science
Abstract: Alzheimer's disease (AD) is often assessed through local hippocampal atrophy, although cognitive decline may also involve disrupted global shape organization. We define geometric fragility as impaired global hippocampal organization rather than an accumulation of local surface changes. Using geometry-only point clouds, we compared local feature aggregation with global hierarchical attention as probes of structural perception. The global model showed higher diagnostic sensitivity and retained recognition under geometric sparsification, whereas the local model showed pattern collapse. Global modeling produced a more ordered disease manifold and a geometric degradation axis that correlated with clinical cognitive decline. These findings suggest that AD progression may involve disruption of global hippocampal shape organization and that hierarchical abstraction can support continuous geometric biomarkers of neurodegeneration.

Keywords: Evolutionary Computation
Abstract: This paper proposes DRACO, a dual-residual adaptive ant colony optimization algorithm for global path planning of mobile robots in two-dimensional grid environments. To mitigate blind exploration, slow convergence, local stagnation, and excessive turning in conventional ant colony optimization (ACO), DRACO integrates four coordinated mechanisms. A line-guided Gaussian pheromone initialization strengthens early goal orientation, whereas a dynamic threelevel node selection mechanism screens ineffective candidates. A dual-residual and direction-continuity based state transition rule jointly evaluates immediate detour redundancy, global deviation from the start-target relation, and directional consistency. In addition, dynamic pheromone regulation with elite-reward updating balances pheromone accumulation and exploration. Comparative experiments on 20×20, 30×30, and 50×50 grid maps and a reverse progressive mechanism validation study evaluate path length, turn count, convergence behavior, and stability. Under the reported experimental settings, DRACO achieves competitive path lengths, favorable turn control, and stable convergence behavior relative to conventional ACO and the tested improved variants.

Keywords: Generative AI, Neural Networks and their Applications, Deep Learning
Abstract: Abstract— Recent research has applied diffusion models to time series forecasting, yet challenges remain in effectively capturing the intrinsic correlations within long multivariate time series forecasting. Modeling the complex dependencies among temporal variables is challenging, as they can be further weakened by the noise inherent in diffusion models. To address these challenges, we propose a Frequency-Domain Dual-Branch Diffusion model (FDBD) for multivariate time series forecasting. Spanning both temporal and spectral domains, FDBD effectively integrates the extraction of multiscale dependencies with the modeling of intricate cross-variable relationships. This dual-domain approach enhances forecasting accuracy and improves the model’s generalization capability. The model leverages the Discrete Wavelet Transform to enable efficient conversion between the time and frequency domains, and employs a dual-branch architecture to separately model the trend and residual components. These dual-branch features serve as conditional inputs for a diffusion model to generate predictions. Experiments conducted on six public datasets demonstrate that FDBD achieves superior overall performance compared with several state-of-the-art time series diffusion models and strong baseline methods. Ablation studies further validate the effectiveness of the decomposition strategy and architectural design. Overall, FDBD offers a robust and scalable solution for multivariate time series forecasting. Keywords— Diffusion model, Discrete Wavelet Transform, Multivariate time series forecasting, Dual-branch architecture, Trend–residual decomposition

Keywords: Generative AI, Expert and Knowledge-Based Systems, Knowledge Acquisition
Abstract: Current Retrieval-Augmented Generation (RAG) systems are typically optimized in silos, struggling to seamlessly unify complex multi-hop reasoning with broad holistic summarization. Furthermore, when agents navigate external knowledge, they often face search path hallucination—proposing unverified queries that cause retrieval failures. To bridge this gap, we propose UniGraphRAG, a unified Agentic Graph RAG framework. During the offline phase, UniGraphRAG employs a threshold-based consolidation mechanism to compress scattered document chunks into dense Node Profiles. Crucially, these profiles explicitly encapsulate both a global semantic summary of the entity and its immediate topological neighbors. During online inference, an LLM metacontroller leverages these dense profiles to dynamically adapt its strategy based on task intent: performing immediate synthesis for global summarization queries, or executing deterministic step-by-step routing guided by the embedded neighbors, which effectively mitigates blind search hallucinations. Extensive experiments demonstrate that UniGraphRAG successfully unifies these dual capabilities, establishing state-of-the-art performance on complex multi-hop QA benchmarks while maintaining highly competitive results in holistic summarization tasks.

Keywords: Reinforcement Learning, Neural Networks and their Applications, Deep Learning
Abstract: This paper introduces the intrinsic curiosity and attention mechanisms into the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, proposing the IA-MADDPG algorithm to address issues such as slow decision-making and weak adaptability in unmanned aerial vehicle (UAV) swarms surrounding maneuvering targets in uncertain environments. First, to address the challenges of UAV swarm cooperative roundup in uncertain environments, we propose a novel framework that leverages the MADDPG algorithm and an intrinsic curiosity incentive mechanism. The MADDPG algorithm enables autonomous control of the UAV swarm, while the curiosity-driven consistency reward model fosters exploration of the unknown environment. In order to expedite the completion of the roundup task, we incorporate a self-attention mechanism into our framework. This mechanism aggregates the state and action information of the UAV swarm, assigning weights to the internal rewards. Simulation results show that the IA-MADDPG algorithm effectively solves environmental uncertainty and reward sparsity in UAV cooperative hunting, achieving faster convergence, higher task completion, and better autonomous hunting performance than MADDPG.

Keywords: Generative AI, Fault Monitoring and Diagnosis, AI and Applications
Abstract: Effective software bug diagnosis requires two ingre dients: the right knowledge source (operational failure history, not just system documentation) and the right retrieval structure (structured records, not unstructured chunks). Current retrieval augmented generation (RAG) approaches fall short on one or both dimensions. We propose Operational Memory RAG (OM-RAG), which indexes resolved issues as structured symptom–root cause–resolution triples and retrieves the most similar historical precedent via single-hop embedding. OM-RAG powers a purpose built large language model (LLM) administrator that has operated a production Dataverse instance for over six months, operationalizing the Knowledge component of the Monitor–Analyze–Plan-Execute–Knowledge (MAPE-K) feedback loop with the system’s own resolved failure history. In a controlled four-configuration ablation, LLM-based judging covers all 1,172 issues and manual verification a 123-issue subset. OM-RAG achieves Diagnosis Accuracy of 0.931 and Fix Correctness of 0.809, outperforming chunk-based retrieval (+186%) and a documentation-anchored concept graph (+77%). Model-free retrieval metrics (mean similarity 0.880) independently corroborate the ranking.

Keywords: Quality and Reliability Engineering, Machine Learning, Deep Learning
Abstract: Just-in-Time Defect Prediction and Localization (JIT-DP and JIT-DL) aim to identify defect-inducing commits and locate defective lines in software systems. However, existing approaches often fail to effectively capture commit semantics or model the interaction between developer intent and code changes, leading to suboptimal prediction and localization performance.

Recent advances in large language models (LLMs) provide new opportunities for improving semantic representations in JIT tasks. However, directly applying LLMs to JIT defect prediction and localization remains challenging due to unstable performance and high inference cost of direct LLM prediction. To address these limitations, we propose JIT-AGM, a unified framework that leverages LLMs for semantic augmentation rather than standalone prediction. JIT-AGM generates structured commit messages and integrates them with CodeBERT-based modeling and a guided attention mechanism to explicitly align commit messages with code changes.

Experiments on the JIT-Defects4J dataset show that JIT-AGM consistently outperforms strong baselines, improving the F1-score from 0.491 to 0.667 and reducing IFA from 10.3 to 6.0, while also achieving gains in effort-aware metrics. These results highlight that LLMs are more effective as semantic augmentation components in our setting, providing new insights into integrating LLMs with structured models for defect prediction and localization.

Keywords: Cognitive Computing, Knowledge Acquisition, Generative AI
Abstract: Humans monitor discourse for referential breaks and repair them by retrieving missing antecedents. Retrieval-based QA systems face a similar problem when chunk boundaries sever coreference chains, leaving otherwise relevant evidence locally underspecified. We propose Monitor–Repair–Integrate (MRI), a cognitively inspired framework for coherence-aware retrieval that detects referential discontinuities, repairs them via targeted retrieval, and integrates the result for generation. We instantiate MRI in MDR-RAG. Human evaluation (N=15) shows a strong preference for repaired snippets (95.4%), and process analyses indicate that performance gains track coherence disruption and repair success rather than evidence volume alone. Across three long-form QA benchmarks and four LLM backbones, MDR-RAG consistently outperforms strong baselines while introducing only 0.34% additional tokens and 3.55s latency per query. These results support MRI as a principled, efficient framework for coherence-aware evidence construction in retrieval-augmented generation. Code is available at: https://github.com/xxy33/MDR-RAG.

Keywords: Cognitive Computing, AI and Applications, Expert and Knowledge-Based Systems
Abstract: Evaluating scientific papers requires identifying functional roles of claims and assessing logical coherence---processes well-studied in expert cognition. We present GraphEval-GT, a framework inspired by expert reasoning that represents papers as structured argumentation graphs with epistemic role assignments. We introduce two attention biases motivated by cognitive research: a position bias reflecting locality constraints in information integration, and a role bias encoding functional dependencies between argument components. A pilot study with 56 reviewers (24 experts, 32 novices) validates these design choices: expert attention concentrates on Method--Result links, with agreement declining as evidence-claim distance increases. Experiments on three benchmarks demonstrate state-of-the-art performance, outperforming frontier LLMs (GPT-5.4, Claude Sonnet 4.6) and existing graph methods. The framework introduces only 12% parameter overhead over vanilla GAT and processes each abstract in under 0.8s. Learned attention patterns align with human behavior (Pearson r=0.72), providing evidence that cognitively inspired structural biases improve both performance and interpretability. Code is available at: https://github.com/xxy33/grapheval-gt.

Keywords: Deep Learning, System Modeling and Control, Neural Networks and their Applications
Abstract: 目前，流体力学中的深度学习研究 主要专注于二维（2D）流动。前往 加速深度学习技术的部署 在实际工程场景中，扩展至关重要 研究范围涵盖高维复杂三维流动 系统。实现三维非稳态的高效预测 本文提出了一种新的时空动力学 由Kolmogorov-Arnold Networks（KAN）和3D驱动的网络 残余卷积，命名为RKAN-3D。模型的准确性 捕捉流场的多尺度空间拓扑 通过三维残留模块，同时利用 KAN 的 可学习的激活机制以缓解拟合 高频传统网络的瓶颈， 非线性物理特征。此外，为了解决 误差发散的严重挑战，广泛存在于 长期自回归预测，本文采用了 趋势预测策略以建模增量特征 流体演化，从而显著增强了 多步的数值稳定性和长期精度 递归推广。实验表明RKAN-3D结构 在预测三维非稳态方&#

Keywords: Cloud, IoT, and Robotics Integration, Deep Learning, Intelligent Industrial Systems
Abstract: While Transformer architectures excel in Long- Term Time Series Forecasting (LTSF), they often fall into a “mean-reversion trap” on non-stationary industrial data, yield- ing overly smoothed predictions that lose vital high-frequency fluctuations. To address this, we propose WavePatchFormer, a lightweight Patch-Centric Transformer designed to balance computational efficiency and waveform fidelity for Industrial Internet of Things (IIoT) applications. Our architecture features three core mechanisms: a joint embedding module for patch-level feature fusion, a Flatten Multi-Layer Perceptron (MLP) direct decoder to mitigate phase drift, and a joint time-frequency loss to preserve spectral energy. Extensive experiments on eight standard benchmarks demonstrate that our model yields lower error rates on datasets with severe distribution shifts (e.g., ETTh2) and reduces shape-matching distances. Finally, we establish the boundary conditions of this paradigm, noting that while explicit channel mixing benefits tightly coupled sensor networks, it creates information bottlenecks in ultra-high-dimensional IIoT deployments.

Keywords: AI and Applications, Neural Networks and their Applications, Application of Artificial Intelligence
Abstract: We address the problem of multi-modal en- tity alignment (MMEA) for multi-modal knowledge graphs (MMKGs). Existing MMEA methods face two core bottlenecks: semantic entanglement caused by insufficient feature disentan- glement, and the lack of collaborative optimization between decoupling and fusion stages. This paper proposes RCMEA, a robust coupled MMEA framework, which achieves fine-grained semantic disentanglement via double orthogonality constraints, and realizes adaptive robust fusion with consistency-aware cross-modal attention. We further build a bidirectional feedback loop for collaborative optimization of the two core stages. Ex- periments on standard MMEA benchmarks show that RCMEA achieves state-of-the-art performance under all common set- tings, with ablation studies validating the effectiveness of our core components.

Keywords: Deep Learning, Biometrics and Applications, Machine Learning
Abstract: 动态手势识别在先进的人机交互和智能控制系统中起着关键作用。然而，有效解读感官数据仍是一个关键挑战，主要原因是局部特征提取不足以及复杂时间依赖性建模困难。本文提出了一种动态手势识别方法，整合了表面肌电图（sEMG）和惯性测量单元（IMU）信号。具体来说，在多通道原始时间信号经过过滤和标准化后，动态手势序列采用滑动窗口策略构建。随后，开发了一种名为TCNN-BLSTM的混合架构，结合了三尺度卷积神经网络和双向长短期记忆网络，用于联合建模局部多尺度特征和长程时间动态。在广泛使用的公共数据集和新建立的包含10名受试者和6个动态手势类的数据集上进行了大量实验。该方法在各自数据集上实现了97.68%和92.03%的窗口级准&#

Keywords: Deep Learning, Transfer Learning, Machine Learning
Abstract: While Contrastive Language-Image Pre-training (CLIP) demonstrates remarkable zero-shot generalization, adapting it to downstream few-shot tasks remains challenging.Existing methods typically treat visual and textual adaptation independently or rely on static fusion, overlooking the critical fine-grained visual context for text refinement and neglecting instance-level uncertainty during inference. To overcome these challenges, we propose a novel method Uncertainty-aware Multi-scale Graph Tuning (UMG-Tuning), which leverages visual injection and graph propagation to optimize cross modal alignment. First, we propose a Visual-Injected Graph Refinement (VIGR) module to semantically enrich abstract text prototypes with concrete visual contexts, explicitly modeling inter-class relationships. Complementarily, we develop an Adaptive Multi-Scale Fusion (AMF) module to aggregate hierarchical visual features and preserve fine-grained details. Finally, we devise an Uncertainty-Aware Logit Fusion (ULF) mechanism that dynamically calibrates predictions by weighing the confidence of different branches, improving adaptation to ambiguous samples. Experimental results show that our method significantly improves the performance of CLIP-based few-shot learning.

Keywords: Deep Learning
Abstract: In recent years, Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have achieved significant progress in medical image recognition, yet they still face a trade-off between capturing global dependencies and computational efficiency. Convolution is limited by its local receptive field, while self-attention mechanisms incur high computational complexity. To address this, we introduce a Neural Ordinary Differential Equations (NODEs) framework to bridge discrete and continuous representations, and adopt improved integration schemes to overcome numerical robustness and long-term integration errors in medical texture analysis. Building on this motivation, we propose GeoLIM-Net, a novel model based on the Line Integral Method (LIM). The solver first applies an explicit Euler predictor to generate an initial path and capture global context, then uses a nested fourth-order Runge-Kutta (RK4) module to refine derivatives and extract nonlinear features, and finally performs implicit line integration to maintain structural consistency. Experiments on BreastMNIST, PneumoniaMNIST, and OrganSMNIST show that GeoLIM-Net achieves state-of-the-art results.

Keywords: Generative AI
Abstract: Human mobility trajectories are fundamental to urban computing and smart city planning. However, strict privacy regulations limit raw data availability, necessitating high-fidelity trajectory synthesis. Despite recent advances, existing generative methods suffer from two limitations: they decouple spatiotemporal dependencies, causing inconsistencies, and they neglect rich contextual factors, disrupting behavioral coherence. To address these, we propose CoST-Diff, a conditional spatiotemporal diffusion framework. By modeling trajectories as unified sequences of spatiotemporal states, CoST-Diff holistically captures the coupling between location and time. Furthermore, urban mobility contexts are incorporated as conditional guidance to steer generation. Extensive experiments on three real-world datasets show that CoST-Diff substantially outperforms state-of-the-art baselines in spatial and temporal fidelity.

Keywords: Deep Learning, Application of Artificial Intelligence, Neural Networks and their Applications
Abstract: Document-level Chinese named entity recognition (CNER) requires contextual evidence at multiple scales, yet common single-receptive-field or fixed-fusion approaches treat token distances symmetrically and do not explicitly model two task priors: boundary tags can depend on asymmetric left/right context, and intra-span attention should decay more gently than cross-span attention. This paper presents DPA-NER, a three-tier pyramid attention framework with a Directional Span-Adaptive Distance (DSAD) bias that encodes these priors through independent directional decay rates and a piecewise decay kernel initialized from dataset span statistics. A dynamic layer weighting (DLW) module fuses per-tier representations for each input, and a hybrid RoBERTa-BiLSTM encoder provides boundary-sensitive embeddings. Experiments on four Chinese NER datasets (QD-NER, Weibo, MSRA, Resume) show that DPA-NER obtains the best F1 among the compared systems on three datasets and ranks second on Weibo.

Keywords: Machine Vision, Image Processing and Pattern Recognition, Application of Artificial Intelligence
Abstract: A joint scene-aware retrieval and scene-classification framework is proposed for e-commerce settings where visually similar products must be separated into fine-grained lifestyle sub-domains. Three key contributions are introduced: (1) a background-agnostic VLM annotation strategy that acquires scene awareness entirely through the training objective, avoiding text-side label injection and shortcut matching; (2) a shared-representation design that co-optimises a symmetric I2T/T2I contrastive loss and a scene head over the same projected image pathway; and (3) gate-controlled cross-block LoRA propagation that links all Q-Former attention blocks under a 4.12% trainable parameter budget, enabling scene-discriminative cues to accumulate across the full attention stack. On a benchmark derived from DeepFashion2 with seven scene categories, the framework achieves a scene classification accuracy of 88.92%, a retrieval R@1 of 53.25%, and a median rank of 1.5.

Keywords: AI and Applications, Connected and Automated Vehicles
Abstract: Existing braking comfort assessment methods for automated vehicles typically classify braking events into broad categories but cannot distinguish between maneuvers that fall within the same category, even when the underlying kinematics differ significantly. This study presents a proof of concept for whether Large Language Models (LLMs) can provide continuous scoring that aligns with established comfort-related indices while offering finer differentiation. Braking data were collected from 91 manual driving runs at a stop-controlled intersection in 31 speed–distance conditions and used as representative braking profiles for evaluating the proposed LLM-based comfort scoring methodology. In phase 1, ten representative runs spanning three ISO 2631 comfort categories were evaluated using five prompting strategies: zero-shot, domain context, few-shot calibration, chain-of-thought, and raw time-series input. Each strategy produced a comfort score on a 1–10 scale, a categorical label, and a natural-language justification. All five strategies achieved strong inverse rank correlation with physical severity (Kendall’s τ = −0.867) and perfect inter-strategy agreement (Kendall’s W = 1.000). In phase 2, the chain-of-thought strategy was applied to all 91 runs. The resulting scores ranged from 1.8 to 8.6, spreading the 65 runs assigned to a single ISO 2631-based comfort category over a range of 1.8 to 6.1. These findings suggest that LLMs can complement established comfort indices by providing continuous and interpretable scoring outputs where categorical methods reach their resolution limit.

Keywords: Decision Support Systems, Machine Learning
Abstract: Volatility forecasting becomes challenging when market conditions shift and model performance varies across market states. Motivated by this instability, we develop a risk-sensitive specialist routing framework for ETF volatility forecasting. The framework uses online risk-sensitive evaluation and state-dependent gating to combine different forecasting specialists across calm and stressed market states. Using a daily panel of six ETFs under a rolling walk-forward design, we find that the strongest forecaster is regime-dependent rather than stable across all states. Relative to the rolling-best baseline, the proposed routing framework reduces high-volatility forecast loss by about 24% and underprediction loss by about 22%. These results suggest that specialist routing provides a practical forecasting architecture that adapts to changing market conditions.

Keywords: System Modeling and Control, Systems Safety and Security
Abstract: Hazard analysis is critical to the development and airworthiness certification of airborne systems. However, current methods face two fundamental limitations: manual qualitative derivation struggles to ensure complete scenario coverage, while probabilistic models suffer from state-space explosion when scaling to multi-source concurrent threats. To address these issues, we propose an integrated framework combining automated reasoning and stochastic modeling. This famework integrates Systems-Theoretic Process Analysis for Security (STPA-Sec) with stratified Datalog reasoning to automatically generate a Causal Scenario Tree (CST) covering multi-source threats. Subsequently, a predecessor node aggregation strategy is used to map the CST to a Generalized Stochastic Petri Net (GSPN), enabling steady-state probability quantification of system-level unsafe control actions while circumventing state explosion. A case study of the Flight Management System (FMS) demonstrates that this method reduces the model solution dimension, quantifies steady-state hazard probabilities, and reveals the nonlinear suppression effect of dynamic recovery mechanisms on cyberattacks. The proposed famework provides rigorous quantitative support for resilience architecture optimization and comprehensive airworthiness assessment in complex airborne systems. Index Terms—Hazard analysis, STPA-Sec, Automated reasoning, Causal Scenario Tree, Predecessor node aggregation

Keywords: AIoT, Cloud, IoT, and Robotics Integration, Cyber-physical systems
Abstract: The rapid advancement of Internet of Things (IoT) technology has led to the widespread connection of embedded devices to the Internet. This evolution provides significant convenience for both industrial production and daily life. However, attackers have recently exploited security vulnerabilities in IoT device firmware. They infiltrate and control devices to compromise user privacy. Furthermore, these attacks pose significant threats to national and social security. Fuzzing is an automated software testing technique. It has successfully identified vulnerabilities across diverse software systems. Researchers have applied fuzzing to firmware testing. This approach has uncovered numerous real-world vulnerabilities. It also significantly enhances the efficiency of firmware security testing. Existing IoT firmware fuzzing methods enhance efficiency through state preservation and taint analysis. These methods also address security risks arising from multi-binary interactions. However, they do not measure the distance from cross-binary basic blocks to sensitive basic blocks. This lack of measurement results in inefficient energy allocation for test cases. We propose GFD-FirmFuzz, a graph fusion-based directed fuzzing method for IoT firmware. The method constructs and fuses binary dependency graphs, function call graphs, and control flow graphs. This fusion enables the calculation of basic block-level distances across binaries. GFD-FirmFuzz then allocates energy to test cases based on these distances to enhance fuzzing efficiency. We developed a prototype of GFD-FirmFuzz on top of STAFF. We selected fifteen IoT device firmwares as test subjects. Our evaluation compared GFD-FirmFuzz against STAFF, AFLNet, and TriforceAFL. Results show that GFD-FirmFuzz increases basic block coverage by an average of 5.29%, 3.41%, and 32.63%. It also triggers 7, 25, and 56 additional unique crashes, respectively. Furthermore, GFD-FirmFuzz discovered more crashes related to multi-binary interactions.

Keywords: Robotic Systems, Reinforcement Learning, Environmental Sensing
Abstract: Deep reinforcement learning has become a core technology for addressing complex motion control problems in humanoid robots. To tackle the challenges of low sample efficiency and policy oscillations in autonomous motion control of humanoid robots in unstructured environments, this paper uses the task of passing through a narrow gate as a representative scenario and constructs an end-to-end control framework based on multiple mainstream algorithms, including DQN, SAC, and GAN-SAC. Furthermore, an experience replay mechanism named HQE-PSS, which is improved by a Hybrid Quality Evaluation and Priority Sampling Strategy, is proposed. Through ablation studies conducted in a Webots-based physical simulation environment for gate passing, the proposed mechanism is shown to effectively resolve the learning failure issue of traditional experience replay under imbalanced sample distributions. It not only improves the convergence speed and gate-passing learning policy of the three baseline algorithms but also significantly enhances the environmental adaptability of humanoid robots in scenarios with randomly dynamic gates. Experimental results demonstrate that the proposed mechanism accelerates algorithm convergence and significantly improves task success rates.

Keywords: Decision Support Systems, Fuzzy Systems and their applications, Systems Safety and Security
Abstract: Emergency decision-making under uncertainty is a research hotspot. In particular, in human-centered real-world scenarios, emergency decision-making needs to balance safety constraints with practical decision needs. However, existing methods obtaining optimal plans by approximating the ideal solution tend to breach safety thresholds by prioritizing resource efficiency; whereas methods for obtaining decision plans through expert-references based often sacrifice optimal resource efficiency due to excessive conservatism in determining attribute weights based on expert experience; furthermore data-driven decision models exhibit insufficient reliability when making decisions within extreme casualty ranges, primarily due to the scarcity of extreme casualty samples. To address these limitations, we propose an emergency multi-attribute decision-making model combining CRITIC-TOPSIS and hierarchical non-orthogonal rule-based Mamdani fuzzy systems (CTF-MADM). The proposed model is implemented through a two-layer architecture: the lower layer utilizes a hybrid approach integrating Criteria Importance Through Intercriteria Correlation (CRITIC) for calculating the objective weight of criteria and the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) for ranking candidate plans to evaluate the resource efficiency of plans. The upper layer employs a hierarchical non-orthogonal rule-based Mamdani fuzzy system, converting domain-specific rules into membership degrees during the fuzzification step. Through fuzzification, the hierarchical non-orthogonal rule base, reasoning system, and defuzzification processing, the system ultimately combines the defuzzified scores with those derived from the lower-level decision plans to derive the final decision plan. Comparative experiments conducted across various fire sizes and multiple fire scenarios demonstrate that the proposed model generates decision plans capable of minimizing casualties in large-scale fire scenarios while maintaining optimal resource efficiency in small-scale fire scenarios.

Keywords: Human-Computer Interaction, Human-centered Learning, Deep Learning
Abstract: Stochastic human motion prediction (SHMP) focuses on forecasting multiple plausible future human pose sequences from past observations. Despite notable advancements, most previous works rely on intricate architectures like RNNs and Transformers, typically necessitating cumbersome training processes and complex loss designs. In this paper, we challenge this convention by introducing MotionMLP, a single-stage and end-to-end MLP-based generative framework for this task. We show that, after integrating well-established practices, a simple MLP-based architecture can significantly outperform current leading approaches. Moreover, to address the inference inefficiency, we further introduce flow matching algorithm to bridge the gap between high performance and low latency. Comprehensive evaluations on Human3.6M and AMASS datasets validate the remarkable superiority of our method.

Keywords: Intelligent Transportation Systems, Reinforcement Learning, Distributed Intelligent Systems
Abstract: Effective coordination in Multi-Agent Traffic Signal Control (M-TSC) relies on efficient inter-agent communication. Existing paradigms face a ``bandwidth-foresight dilemma'': sharing raw hidden states is communication-efficient but lacks strategic forward-looking guidance, while sharing explicit future traffic predictions incurs high computational and bandwidth overheads. To circumvent this, Intention Broadcasting presents a compact alternative where agents share their anticipated action-value distributions. However, generating reliable intentions requires deeply understanding dynamic traffic evolution rather than relying on static, single-step observations. To address this limitation, we propose STGIC, a novel framework that integrates temporal foresight with spatial coordination. Specifically, STGIC employs a sliding-window temporal encoder to process short-term historical observations, implicitly extracting local traffic trends to generate trend-aware intentions. Subsequently, a relational transform graph attention network routes these intentions spatially, dynamically filtering neighborhood noise based on local temporal states. Extensive evaluations on real-world datasets demonstrate that STGIC achieves highly competitive performance against representative baselines. Furthermore, mechanistic ablations validate the indispensability of coupling temporal trend extraction with spatial intention sharing.

Keywords: Quality and Reliability Engineering, Deep Learning, Systems Safety and Security
Abstract: During the development and maintenance of software systems, the introduction of faults is inevitable, posing significant threats to software reliability and security. To accurately locate these faults, researchers have developed various automated fault localization methods. These methods include spectrum-based, mutation-based, and learning to rank-based approaches, which primarily analyze program execution information and code features to locate faults. However, these methods still face limitations in understanding the semantic information of faults, which limits their accuracy of fault localization. To address the challenge, this paper proposes LLMPAFL, which is a large language model prompt augmentation-based statement-level software fault localization. Firstly, the method utilizes a deep learning model to assign suspicious scores to code statements of the program, and select the methods containing highly suspicious statements to construct the code context. This approach alleviates the context window constraints encountered by large language model (LLM) when handling large-scale projects. Secondly, LLMPAFL integrates multi-source information, including bug report, stack trace information and suspicious scores, to construct structured prompts that guide LLM in performing multi-stage reasoning. Finally, it outputs a ranked list of suspicious statements to achieve precise fault localization. Experimental results demonstrate that when examining the first statement in the ranked list of suspicious statements, LLMPAFL successfully locates 19-82 faulty statements compared to spectrum-based, mutation-based, and learning to rank-based methods in fault localization.

Keywords: AI and Applications, Image Processing and Pattern Recognition, Deep Learning
Abstract: 无人机的部署 （无人机）在军事侦察中带来了极高的要求 关于 他们的机载物体检测系统用以识别微小的， 密集 复杂环境中的目标。为了解决召回率低的问题 对小型目标的探测率和严重漏接率 此类场景下的通用探测器，本文 提出一个 高精度分析增强型探测器——IMSFC-YOLO。 基于YOLOv11的基础，核心创新 包括：1）设计内部多尺度小靶 增强模块（IMSEM），加强了这些功能 以及通过通道分割实现微小目标的轮廓 多尺度 池化和拉普拉斯边缘提取;2）构造一个 轻量级功能上下文模块（FCM）高效地 通过 聚合多尺度上下文信息 平行 多路径结构;3）引入浅层检测 头部 利用高分辨率特征直接改进 Smalltarget召回。VisDrone2019数据集上的实验 节目 IMSFC-YOLO实现47.1%的mAP@0.5%，并实现召回 45.3%的感染率，É

Keywords: Deep Learning, Machine Learning, Representation Learning
Abstract: Multivariate time series often exhibit complex temporal variation patterns and inter-channel correlations. These two factors are closely related, yet most existing methods fail to jointly exploit them. Temporal variations typically show multi-scale features, whereas inter-channel correlations often encompass both commonalities and distinct characteristics across channels. Motivated by this observation and to address this limitation, we design dedicated mechanisms for the temporal and channel dimensions and propose D2Net, whose core consists of the Multi-Scale Progressive Mixing module (MSPM) and the Channel Common–Specific Mixing module (CSM). Firstly, MSPM fuses multi-scale temporal information, followed by CSM, which separates and fuses channel common and channel specific patterns. To capture global trends and local fluctuations, D2Net then adopts a dual branch network consisting of a linear branch and a non-linear branch, and adaptively fuses the outputs of the two branches for the final prediction. Finally, extensive comparative and ablation experiments on nine real-world datasets validate the strong performance and efficiency of D2Net.

Keywords: Generative AI, Machine Vision, Cognitive Computing
Abstract: 在具身人工智能中部署视觉语言模型需要 但仍然具备稳健的同种空间推理 跨参考系的透视取向依然存在 对现有车型来说颇具挑战性。我们假设那是 贡献因素是参考系挑战 标准空间推理，直接推理严格 将以镜头为中心的视觉感知与隐含的 协调变换。为了缓解这种情况 表征纠缠，我们提出级联空间纠缠 提示（CSP），一种推理时间提示策略，具有 将异心空间关系预测分解为两部分 显式阶段：（i） 原子摄像机视角感知 （提取宾语关系和人称取向）以及 （ii） 参考框架的规则引导文本推理 转变。系统性评估此类多步骤 基于这个原因，我们引入了AlloCam-Set，一种诊断工具 评估集仅包含约3000人 对齐的三阶段层级注释。广泛 实验显示CSP显著改善了平均值 AlloR

Keywords: Application of Artificial Intelligence, Deep Learning, Smart Buildings, Smart Cities and Infrastructures
Abstract: Abstract— Shield tunneling is the predominant method for complex urban subterranean construction. Maintaining the precise attitude of the shield machine is critical to prevent structural damage and ground settlement. However, unpredictable geological variations induce sudden attitude deviations, requiring accurate predictive models for proactive correction. While deep learning is increasingly employed, current data-driven approaches exhibit two limitations: they disregard rigid-body kinematics, leading to inconsistent spatial predictions, and output deterministic point estimates, lacking probabilistic bounds for robust risk assessment.

This paper proposes a physics-guided, dual-paradigm architecture synergizing a hybrid LSTM-Transformer network with mechanics-aware feature engineering. A Rigid-Body Constraint Loss (RBCL) is embedded to enforce kinematic consistency, and quantile regression maps deterministic estimates into calibrated uncertainty intervals. Validation on 1,647 excavation rings from the Huanggang Cross-River Tunnel reveals task-dependent advantages. A pure Transformer achieves the lowest mean absolute error (MAE) of 2.680 mm in deterministic point prediction (Paradigm I). Conversely, the hybrid LSTM-Transformer demonstrates exceptional robustness in uncertainty quantification (Paradigm II), yielding a precise MAE of 2.742 mm and an 80% Prediction Interval (PI) with 82.13% coverage and a 9.20 mm width. Gradient saliency and SHAP analyses elucidate the spatiotemporal learning mechanisms. By balancing deterministic precision and calibrated uncertainty, this framework facilitates proactive, mechanics-informed risk management.

Index Terms— Tunnel, Shield Machine, Attitude Prediction, Rigid-Body Constraint, Prediction Interval (PI), LSTM-Transformer.

Keywords: Systems Safety and Security, Generative AI, Information Assurance and Intelligence
Abstract: Model Context Protocol (MCP) has emerged as the de facto interoperability layer for connecting LLM agents to external tools. In project settings, however, repository-distributed MCP configuration files increasingly function not only as capability declarations, but also as portable, approvable, and executable local startup entry points. Prior work has studied protocol threats, platform benchmarks, and metadata poisoning in MCP ecosystems, yet the access-time risk of executable project-level MCP configurations remains under-modeled. We present MCP-Guard, a layered framework that combines rule-based prefiltering with LLM-based semantic analysis to detect high-risk startup semantics in project-level MCP configurations, and a pre-execution policy guard for host-side onboarding. The detector reconstructs execution semantics from configuration files, installation instructions, and auxiliary scripts, and identifies patterns such as shell wrapping, runtime fetching, capability asymmetry, and semantic obfuscation. The guard enforces source binding, configuration-drift constraints, semantic risk explanation, and policy mapping before a local server is launched. On a manually labeled evaluation set, MCP-Guard achieves 89.7% precision, 86.8% recall, and 86.1% macro-F1, outperforming rule-only and lightweight semantic baselines. In pre-execution protection experiments, the guard raises the block rate for high-risk configurations from 8.7% to 86.8% with only 0.83 s average overhead.

Keywords: AI and Applications, Deep Learning, Image Processing and Pattern Recognition
Abstract: With the rapid advancement of deep generative technologies, multimedia such as audio and videos has become inundated with extensive forged material, posing significant challenges to the security of communication content. Furthermore, the rising incidence of telecom fraud cases underscores violations of biometric privacy, such as the misuse of human voices and facial features. Notably, forged videos synthesized through lip synchronization are realistic. Existing detection methods typically rely on identifying audio-visual out-synchronization to differentiate genuine content from fake, yet their performance remains largely unsatisfactory. In this paper, we propose a multimodal lip-sync forgery detector based on multi-scale attention and orthogonal fusion. On the one hand, our approach captures fine-grained multi-scale spatial information around the lip region and allocates distinct attention weights to prioritize key evidence for authenticity evaluation. On the other hand, it introduces an orthogonal fusion mechanism to enhance the interaction between global temporal features and local spatial features, thereby boosting the accuracy of identifying discrepancies between lip movements and corresponding audio signals. Extensive experiments validate the effectiveness and generalizability of our approach, which is designed to mitigate the escalating risks in current communication and multimedia security.

Keywords: System Modeling and Control, Robotic Systems
Abstract: This study presents a physics-model-based co-optimization framework for the design and locomotion of a three-link robotic fish based on the Sliding Strip Discrete Vortex Method and Bayesian optimization. By incorporating a minimum speed constraint, the proposed method avoids degenerate low-speed solutions and enables efficient identification of physically meaningful designs under limited function evaluations. This constraint effectively reshapes the optimization landscape by penalizing insufficient propulsion, thereby guiding the search toward regions that balance energy efficiency and swimming performance. The optimization achieves a minimum cost of transport of 16.707 J/(kg*m). Results show that the design space exhibits a structured near-optimal region, where different optimization runs converge to comparable performance despite variations in optimal design variables. The proposed framework provides an efficient and consistent method for coupled morphology-kinematics optimization of bio-inspired robotic fish.

Keywords: Fault Monitoring and Diagnosis, Deep Learning, Artificial Social Intelligence
Abstract: 跨不同工业领域 0340;轴承故障诊断 工作条件要求并 4892;的框架 可概括、判别且 1487;解释， 还没 现有方法难以统
9968;异质振动 来自不同传感器 0340;信号，提取全ह ;6;光谱 德彭 频域表示的密度 5292;产生 超越离散故障标 1614;的可操作诊断๛ ;5;出。我们 建议Trans-BearLLM，一位ੋ ;6;强健的健康男性 引入了变形金刚 2686;强的年龄框架 故障分类网络以 1462;代原有的 多尺度 卷积注意力块与 1472;加自我注意力 赛段 以及跨频下采样 5292;实现全局频谱 依赖关系 在 由离散余弦归一 1270;产生。该 提取 振动特征被对齐 1040;的令牌空间中 Qwen2 通过语义初始化 0340;对齐层和 好吧 通过LoRA调校，使模 ;型能够同时实现 进行故障诊断、 4322;常检测、维护ॿ ;4;议 通过自然语言进 4892;豁免与风险分ਜ਼ ;2;报告生成。MBHM数据 ;集上的实&

Keywords: Generative AI, Deep Learning, Application of Artificial Intelligence
Abstract: Decompilation aims to recover more understandable high-level code from low-level binaries and is a fundamental capability in security tasks such as vulnerability discovery, malware analysis, and program understanding. In recent years, large language models have substantially advanced neural decompilation. However, existing methods mainly focus on model architectures, training scale, or inference strategies, while the quality variation of supervision targets remains underexplored. To address this issue, this paper proposes MUSE-Dec, a multi-aspect supervision evaluation framework for neural C++ decompilation. The proposed method first transforms raw pseudo-C into more source-like supervision targets through a two-phase lifting process, and then evaluates the target from four aspects, namely syntactic validity, diagnostic cleanliness, decompilation artifact burden, and structural readability. Based on the resulting quality scores, the training data are further filtered, tiered, and resampled to construct a higher-quality training distribution. Experimental results on three complementary benchmarks show that MUSE-Dec achieves the best results on HumanEval-Decompile and GitHub2025, while remaining highly competitive on BinMetric-DEC and improving the backbone by 21.7%. These results indicate that explicitly modeling supervision target quality can significantly improve neural decompilation performance without increasing model size.

Keywords: Biometrics and Applications, Machine Vision
Abstract: Cranial nerve dysfunction manifests in diverse clinical conditions, of which facial paralysis is among the most prevalent and clinically significant. However, its clinical diagnosis still depends largely on subjective visual observation, which limits consistency and reproducibility in practice. Despite recent advances in computer-aided diagnosis, existing methods often fail to jointly ensure feature stability, clinical interpretability, and diagnostic robustness. In this paper, we propose CAMEL-FP, a hierarchical framework for interpretable cranial nerve health diagnosis based on Clinically-guided Action-level Modeling and Explainable Learning, using facial paralysis as a representative disease scenario. Specifically, CAMEL-FP represents facial dysfunction using structured indicators derived from facial landmarks, including symmetry, angular variation, and motion velocity differences. A significance-guided strategy is used to adaptively identify discriminative indicators for each action, and action-specific models equipped with clinically informed prior constraints are employed to capture diagnostic patterns while preserving consistency with medical knowledge. The resulting action-level evidence is then aggregated through an adaptive fusion module to form a global diagnostic representation, with feature contribution analysis providing additional interpretability. Experimental results on a clinical dataset show that CAMEL-FP achieves an AUC of 0.9904 pm 0.0086, a sensitivity of 0.9580 pm 0.0390, a specificity of 1.0000 pm 0.0000, and an F1-score of 0.9780 pm 0.0210, indicating its strong effectiveness, stability, and potential clinical utility.

Keywords: Intelligent Transportation Systems, Cyber-physical systems, Modeling of Autonomous Systems
Abstract: Federated Learning (FL) in V2X vehicular environments operates under varying connectivity, computational availability, and node participation, making it difficult to evaluate its temporal behavior across runs. Approaches based exclusively on simulation require multiple runs and do not allow for the controlled isolation of the effects of concurrency and synchronization on system performance. This work proposes a stochastic analytical model to represent the execution cycle and temporal behavior of FL in 5G/6G vehicular architectures. Validation against experiments implemented in the Flower framework showed differences of less than 5% in average round time, with mean values of 28.77s for the SPN model and 29.31s in the experimental environment. The performance analysis also showed that round time increases from about 7 s to 53 s, while round rate decreases from 0.14 to 0.021 rounds/s as training time and the number of clients grow, highlighting the effects of synchronization and processing overhead on system behavior.

Keywords: Medical Informatics, Image Processing and Pattern Recognition, Deep Learning
Abstract: Semi-supervised medical image segmentation effectively reduces annotation costs by leveraging limited labels alongside abundant unlabeled data. However, existing methods often fail to fully exploit structural priors in labeled data or maintain stable class-wise semantic references across iterations, making unlabeled feature learning prone to representation drift and noisy pseudo-label propagation. To address this, we propose CPPMC-Net, a framework that integrates correlation-aware priors with prototype memory consistency. Built on a student-teacher-student mutual learning framework, CPPMC-Net employs a Prior-Guided Information Extraction (PIE) module to model cross-sample interactions, enabling the transfer of structural anchors from labeled to unlabeled data. Furthermore, a Prototype Memory Consistency (PMC) module serves as a global state anchor, maintaining stable class-wise references across iterations and enforcing prototype-relation consistency to suppress noise propagation. Experiments on the ISIC and LA datasets demonstrate that CPPMC-Net achieves superior performance, particularly in boundary accuracy and structural stability, highlighting its potential for reliable medical image analysis in cyber-medical applications.

Keywords: Decision Support Systems, Quality and Reliability Engineering, Fault Monitoring and Diagnosis
Abstract: Daily ETF risk monitoring can become unreliable when market data quality degrades, market conditions shift, or predictive performance becomes unstable. This paper develops a reliability-aware risk monitoring service for next-day tail-risk surveillance. The proposed framework combines service-time quality checks, lower-tail prediction, uncertainty scoring, and risk-aware adjustment of the tail-risk estimate. We evaluate the system on a daily panel of multiple ETFs augmented with VIX and yield-curve information under a rolling walk-forward design. Empirically, the framework improves tail-risk monitoring, especially during stressed periods, while remaining reliable under simulated input degradation.

Keywords: Metaheuristic Algorithms, Heuristic Algorithms, Machine Learning
Abstract: In ultra-high-definition (4K/8K) video transmission, such as remote surgeries and augmented reality, latency and uplink limitations pose challenges for agile provisioning. Ensuring QoS in these critical applications demands intelligent, low-latency network management. The 5G system prepares the environment in advance through network slicing. Mapping these slices is a computationally complex (NP-hard) problem, traditionally solved using heuristics that require significant time. This work proposes a hybrid approach that integrates an ML-based admission control mechanism into the heuristic mapping process. The solution uses predictive modeling to assess request feasibility before mapping, considering 13 QoS parameters and the virtual network's characteristics. Requests with a high probability of failure or SLA (Service Level Agreement) violation are discarded in advance, reducing customer response time and optimizing the use of the provider's computational resources. The results indicate that predictive filtering accelerates provisioning and increases the mapping success rate for high-demand video services; in the best case, the time dropped from 41.93s to 0.06s, over a 1,000x reduction, allowing real-time operation.

Keywords: Cyber Systems and Engineering, Resilience Engineering, Cyber-physical systems
Abstract: The increasing interconnectivity of digital systems has expanded the attack surface of critical infrastructures, requiring more adaptive defenses. The Moving Target Defense (MTD) paradigm emerges as an alternative solution by introducing unpredictability into system configurations. This paper proposes a performance evaluation model based on Stochastic Petri Nets (SPNs) to analyze the MTD method from a Diversity-of-Resources perspective, accounting for recognition time, success probability, and resilience. Results indicate that increasing resource diversity can raise the Mean Time To Absorption (MTTA) by up to 357% and significantly reduce the attacker's success rate, demonstrating that periodic heterogeneity changes the attack-defense dynamic. Under concurrent attack scenarios, MTTA with diversity reaches up to ten times the value observed in static environments, further underscoring its effectiveness under high-concurrency conditions.

Keywords: Biometrics and Applications, Biometric Systems and Bioinformatics, AI and Applications
Abstract: Deploying deep learning models for real-time electrocardiogram (ECG) analysis on wearable and edge devices is challenging due to their high computational cost and large parameter size. To address this issue, we propose LiteECG-Net, a lightweight yet effective network for ECG arrhythmia classification. LiteECG-Net integrates a dual-branch frequency-temporal feature extractor, multi-scale depthwise separable convolutions, and a channel-temporal attention module to learn discriminative ECG representations with low computational overhead. Experiments on the PTB Diagnostic ECG Database and the MIT-BIH Arrhythmia Database show that LiteECG-Net achieves 99.7% accuracy on PTB and 99.3% accuracy on the stratified MIT-BIH split, while obtaining the best F1 score on the more challenging patient-wise split. Compared with lightweight baselines, LiteECG-Net reduces the parameter count by 31.8% and the model size by 68.9%. Moreover, hardware benchmarking demonstrates its practical edge-deployment capability, achieving an inference latency of 3.8 ms on NVIDIA Jetson Orin and 9.4 ms on Raspberry Pi 4B. The dataset and code are available at https://github.com/houchenyu/LiteECG.

Keywords: Reinforcement Learning, Neural Networks and their Applications, Human-Machine Interaction
Abstract: Multi-degree-of-freedom robotic arms are widely used in unstructured dynamic environments such as collaborative services and agricultural harvesting. Achieving efficient and safe path planning in complex environments is a prerequisite for completing high-precision tasks. However, compared to traditional structured environments, the high-frequency interaction between robots and environment leads to significant uncertainty. Furthermore, the high-dimensional motion space of robotic arms results in immense computational complexity, making real-time reactive planning a challenge. Therefore, we propose a hierarchical planning method based on global path guidance to achieve a balance between computational efficiency and flexibility. Specifically, the global planning module first generates a collision-free reference path in the configuration space at the initial moment when obstacles are stationary, ensuring that the target configuration can be reached. Then the reference path is decomposed into continuous local target points and added as local constraints to the optimization objective of reinforcement learning. Finally, the reinforcement learning module is used as a local planner to generate reactive actions based on robot state and environment information. Simulation results show that our method outperforms standard off-policy reinforcement learning algorithms in terms of learning speed and arrival accuracy.

Keywords: Robotic Systems, Environmental Sensing, Deep Learning
Abstract: Achieving safe, high-speed autonomous flight in complex environments with static, dynamic, or mixed obstacles remains challenging, as a single perception modality is incomplete. Depth cameras are effective for static objects but suffer from motion blur at high speeds. Conversely, event cameras excel at capturing rapid motion but struggle to perceive static scenes. To exploit the complementary strengths of both sensors, we propose an end-to-end flight control network that achieves feature-level fusion of depth images and event data through a bidirectional crossattention module. The end-to-end network is trained via imitation learning, which relies on high-quality supervision. Building on this insight, we design an efficient expert planner using Spherical Principal Search (SPS). This planner reduces computational complexity from O(n^2) to O(n) while generating smoother trajectories, achieving over 80% success rate at 17 m/s—nearly 20% higher than traditional planners. Simulation experiments show that our method attains a 70–80% success rate at 17 m/s across varied scenes, surpassing single-modality and unidirectional fusion models by 10–20%. These results demonstrate that bidirectional fusion effectively integrates event and depth information, enabling more reliable obstacle avoidance in complex environments with both static and dynamic objects.

Keywords: AI and Applications, Generative AI
Abstract: Despite the impressive performance of large-scale vision-language pre-training in multimodal tasks, existing models still struggle with fine-grained understanding. As shown in Figure. 1(d), the models fail to differentiate between “a man in the distance wearing a blue shirt” and “a man in the distance wearing a red shirt,” as they tend to focus on prominent cues while overlook fine-grained information. In this paper, we propose FG-CLIP, which leverages the Large Language and Vision Assistant (LLaVA) to enhance multimodal fine-grained representations. We first use LLaVA to generate fine-grained and negative semantics, guiding the model to learn fine-grained features. We then introduce a Fine-Grained Knowledge Enhancement Framework that further refines fine-grained representations through a Dual Text alignment (DTA) strategy and a Fine-Grained Enhancement Encoder (FGE). We introduce two challenging datasets, VGHard-Attribution and VG-Hard-Relation, to assess fine-grained representation ability. Experimental results demonstrate that FG-CLIP achieves state-of-the-art (SOTA) performance on the VG-Attribution, VG-Relation, VG-Hard-Attribution, and VGHard-Relation datasets, surpassing the previous SOTA method by 5.1%, 3.7%, 15.8%, and 12.9%, respectively. Results on general representation tasks indicate that FG-CLIP enhances fine-grained representations while preserving general representations. Additionally, FG-CLIP effectively integrates finegrained textual information into generated images, enabling fine-grained visual generation.

Keywords: Adaptive Systems, Deep Learning, Knowledge Acquisition
Abstract: Retrieval-Augmented Generation (RAG) grounds large language models (LLMs) in external knowledge to mitigate hallucinations, yet indiscriminate retrieval introduces noisy documents and incurs redundant latency. We propose SACG-RAG (Semantic-Adaptive Cluster-Guided RAG), a framework that replaces uniform retrieval with multidimensional semantic analysis. SACG-RAG routes queries to distinct pathways---direct generation, explicit denoising, or cluster-guided tree search---based on intrinsic semantic characteristics including commonsense dependency, temporality, specificity, ambiguity, and precision. For multi-hop reasoning, we employ semantic clustering and consensus-weighted Monte Carlo Tree Search (MCTS) to eliminate redundant exploration. Experiments on TriviaQA, PopQA, Natural Questions, and ASQA demonstrate that SACG-RAG reduces average retrieval calls by 25% compared to Adaptive-RAG while improving accuracy by 6.8% on average across datasets. Our contributions are threefold: (1) a label-free (at deployment) semantic routing mechanism using pre-trained components for query classification; (2) an explicit denoising mechanism for single-step retrieval; and (3) a cluster-guided tree search with confidence-based early stopping for efficient multi-hop reasoning.

Keywords: Deep Learning, Image Processing and Pattern Recognition, Hybrid Models of Neural Networks, Fuzzy Systems, and Evolutionary Computing
Abstract: Deep learning-based single-image super-resolution (SISR) persistently faces a trade-off between computational efficiency and long-range dependency modeling. While Convolutional Neural Networks (CNNs) offer high efficiency, their restricted local receptive fields hinder holistic structural recovery. Conversely, recent State Space Models (SSMs), such as Mamba, effectively capture global dependencies with linear complexity but are bottlenecked by inherent causal masking and the heavy computational redundancy of multi-directional scanning strategies. To break this bottleneck, we propose LKMamba, a streamlined and lightweight hybrid architecture. Deviating from conventional complex hierarchical stacking, LKMamba features an elegant dual-stage processing paradigm. First, an Optimized Large Selective Kernel (OLSK) module is introduced to dynamically expand the effective receptive field, ensuring precise multi-scale local feature extraction. Second, to circumvent the causal limitations of SSMs without resorting to expensive multi-directional scanning, we design a novel State-space Prompt Module (SPM). By efficiently injecting learned prompt-based global priors into a single unidirectional scan, SPM achieves non-causal global perception with minimal parameter overhead. Extensive evaluations across five standard benchmarks validate the superiority of LKMamba. Notably, compared to the highly competitive SwinIR-light, LKMamba achieves a 0.33 dB PSNR improvement on the Manga109 dataset while reducing computational costs (MACs) by 28%, demonstrating a better trade-off between accuracy and efficiency.

Keywords: Medical Informatics, Machine Learning, Deep Learning
Abstract: Accurate and accessible blood pressure (BP) risk stratification remains challenging in real-world settings because conventional cuff-based measurements are intermittent, contact-dependent, and impractical for frequent monitoring. This study presents a fully contactless framework for BP risk stratification using single-site facial video-derived remote photoplethysmography (rPPG) derived from a single-site facial video acquired from a smartphone camera. The proposed approach combines modified UNet-based waveform reconstruction, PPG-based multi-domain feature extraction, and hierarchical ensemble classification to distinguish low, normal, and hypertensive BP categories. The framework is evaluated on the publicly available PulseDB dataset and a curated smartphone dataset collected under structured cardiovascular perturbations. Contact PPG signals from PulseDB are used after preprocessing, while smartphone facial rPPG signals undergo region-of-interest extraction, spatial averaging, filtering, and morphology-preserving reconstruction to recover BP-sensitive waveform characteristics. Temporal, morphological, ratio-based, and demographic features are then extracted to capture complementary cardiovascular information relevant to BP variation. Experimental results show that integrating demographic information with PPG-derived features consistently improves classification performance, with the best-performing model achieving accuracies of 0.79 on PulseDB and 0.91 on the smartphone dataset. The results further indicate that hierarchical classification improves robustness under overlapping class boundaries, while waveform reconstruction enhances the reliability of BP-oriented feature extraction from contactless signals. These findings establish the proposed framework as a practical and physiologically grounded solution for smartphone-based BP risk stratification, with strong potential for future real-time mobile health deployment.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Machine Learning
Abstract: Accurate 3D segmentation of the hippocampus from MRI scans plays an important role in early detection of neurodegenerative diseases, however, precise segmentation is often hindered by small target volumes and complex geometries, which lead to feature degradation in deep network layers. To overcome these issues, we design Structural-Spatial-Gated Net (SSG-Net), which combines multi-scale feature extraction with a gated spatial attention mechanism. The encoder em- ploys the DPC Block, which utilizes 7 × 7 × 7 depthwise convolutions to enlarge the receptive field without sacrificing fine-grained structural details. To resolve morphological am- biguities, the Multi-Scale Feature-Context Calibration Block (MS-CCB) leverages dilated group convolutions for multi-scale perception, subsequently performing context-aware distribution alignment to distill salient hippocampal textures from noisy backgrounds. Furthermore, Gated Spatial-Structural Atten- tion 3D (GSSA3D) is integrated into skip connections to decouple orthogonal spatial dependencies and sharpen blurred boundaries via dynamic gating. Extensive evaluations on MSD and ADNI-HarP datasets demonstrate that SSG-Net shows consistent improvements over several representative methods, with Dice/IoU reaching 93.42%/87.67% and 95.13%/90.71%, respectively. SSG-Net provides a reasonable trade-off between accuracy and computational cost: its computational complexity (GFLOPs) is lower than contemporary state-space models, while its parameter scale significantly outperforms massive Transformer architectures.

Keywords: Neural Networks and their Applications, Deep Learning
Abstract: Regarding intrusion detection in resource-constrained Internet of Things (IoT) environments, knowledge distillation (KD) offers an effective solution for developing lightweight and efficient models. However, existing methods generally lack explainability and are limited to imitating the behavior of teacher models, rather than understanding their internal decision-making logic. Moreover, the selection of distillation data significantly impacts the model performance, while existing research has paid little attention to this aspect. In this paper, we propose an intrusion detection framework called INtrusion Detection based on EXplanation-guided knowledge distillation and coreset selection (INDEX). This framework utilizes attribution explanations as transferable knowledge to explicitly convey the reasoning evidence of the teacher model, thereby enhancing the detection performance and explainability of the student model. Meanwhile, we propose an iterative coreset selection strategy that filters data samples with reliable teacher supervision signals and high informational value for distillation, improving the learning effectiveness of the student. Experimental results on the CIC-IDS-2017, RT-IoT2022, and TON_IoT datasets demonstrate that our proposed INDEX achieves superior detection performance compared to state-of-the-art methods while maintaining lower computational complexity and storage overhead.

Keywords: Deep Learning, Representation Learning, Affective Computing
Abstract: Empathetic dialogue generation has gained significant attention in recent years, aiming to create more human-like and emotionally engaging conversational AI systems. However, existing approaches often struggle to capture the intricate dynamics and nuances of emotional expressions in human communication, leading to generic or contextually inappropriate responses. In this paper, we propose a novel neural information processing approach to empathetic dialogue generation that focuses on optimizing the neural transmission of affective information from the perspectives of language and emotion. Unlike previous methods that neglect the variability of emotional words and the complex interactions between emotions and semantics, our proposed neural emotion-semantics transmission network (NESTN) dynamically captures nuanced emotional expressions and exploits linguistic associations to enhance the model's ability to understand and generate empathetic responses. NESTN incorporates a dynamic emotion-semantics encoder to learn contextually-grounded affective representations and a neural transmission decoder to adaptively control the flow of emotional and semantic information during the generation process. Moreover, we introduce a multi-level attention mechanism to prioritize the most relevant emotional and semantic cues at different granularities. Further analysis reveals that the proposed approach aligns well with psychological theories of empathy, indicating its effectiveness in capturing the core aspects of empathetic communication.

Keywords: Artificial Social Intelligence, Deep Learning, Generative AI
Abstract: Generating high-quality long term multivariate time series is important for mitigating data scarcity in domains such as healthcare, electricity, and energy. Existing generative models are usually designed from a single view, which makes it difficult to jointly model temporal dependencies, inter-variable correlations, and global periodic patterns in long sequences. This limitation reduces generation quality during denoising. To address this issue, we propose TriDiff, a multi-view parallel diffusion model with a Fluctuation-Aware Graph Module and a Dual-Mode Frequency Module. Built on a diffusion framework with an encoder-decoder Transformer and decoupled temporal representations, TriDiff forms a multi-view denoising encoder that fuses temporal, spatial, and frequency-domain features in parallel, improving the use of complementary information across views. Experiments on five public datasets, namely Stocks, Energy, fMRI, Mujoco, and Sines, show that compared with the representative baseline DiffusionTS, TriDiff reduces Context-FID, Correlational Score, and Discriminative Score by 23.63%, 18.84%, and 10.90%, respectively. These results show that TriDiff improves the quality of long term multivariate time series generation.

Keywords: Deep Learning, Generative AI, Transfer Learning
Abstract: Post-training quantization (PTQ) is essential for deploying large language models (LLMs) under hardware constraints. However, low-bit quantization typically suffers from severe performance degradation primarily caused by activation outliers. While recent transformation-based methods attempt to address this, they often incur massive computational overhead during calibration and easily trigger overfitting under limited few-shot data. In this work, we propose DAKQuant, an efficient PTQ framework based on Distribution optimization and Adaptive Knowledge distillation. DAKQuant performs unconstrained orthogonal optimization of rotation matrices via the Whip loss and optimizes scaling matrices through prediction-deviation-based adaptive distillation. By employing Top-K selection to filter long-tail noise, our method successfully suppresses activation outliers while strictly preventing few-shot overfitting. Extensive experiments on LLaMA-2-7B demonstrate that under the extreme W4A4KV4 configuration, DAKQuant effectively retains over 96% of the original FP16 performance. Notably, it completes calibration in just 6.2 minutes on a single GPU, achieving up to a 5.4x speedup over advanced baselines, providing a robust and highly efficient solution for low-bit LLM deployment.

Keywords: Machine Learning, Ethics of AI and Pervasive Systems, Neural Networks and their Applications
Abstract: Data classification underpins critical applications in human society, yet its deployment on sensitive data presents a fundamental dilemma: balancing high accuracy with robust privacy protection that satisfies privacy regulations. Differential Privacy (DP) is the prevailing technical solution, but its standard uniform-noise implementation severely degrades classification accuracy, due to its perturbed discriminative features. Inspired by the dual-process theory and the human selective attention in category learning, we propose a neuro-symbolic model that can adaptively learn features under strict privacy constraints. The policy engine in our model dynamically allocates the privacy budgets based on the feature importance, thereby applying the feature-adaptive noise rather than the uniform noise. In this way, it preserves the informative features while suppressing non-essential ones. Particularly, we formalize regulations (e.g., GDPR) as differentiable symbolic constraints, and thus can seamlessly embed the compliance into the learning process. Our model reduces the accuracy loss significantly under the identical privacy guarantees, demonstrating its effectiveness and competitiveness.

Keywords: Reinforcement Learning, Human-Machine Cooperation and Systems, Robotic Systems
Abstract: Long-horizon robotic tasks remain challenging for reinforcement learning because rewards are sparse and effective action sequences are difficult to discover through flat exploration alone. Hierarchical reinforcement learning mitigates this issue through temporal abstraction, but existing option-based approaches typically learn option boundaries solely from reward signals or depend on manually specified subgoals, which limits semantic interpretability and sample efficiency. We propose emph{VLM-Option}, a hierarchical framework that leverages pretrained foundation models to generate executable skill and context-aware subgoals for low-level reinforcement learning. Given a language instruction and an initial visual observation, a planner selects skills from a predefined library and predicts termination conditions conditioned on downstream task requirements. A goal-conditioned low-level policy is then trained using CLIP-based multimodal representations, where CLIP encodes the active skill description and current image into a shared embedding space. The generated subgoals additionally induce dense intrinsic rewards that complement sparse task feedback. Experiments on Meta-World and Robosuite manipulation tasks show that VLM-Option converges faster, achieves higher success rates, and requires fewer execution steps than E2E (DRQ-v2) RL, HRL (Option-Critic) and PSL (Plan-Seq-Learn) baselines.

Keywords: Deep Learning, Transfer Learning, Representation Learning
Abstract: Prompt-based learning has emerged as an effective paradigm for the parameter-efficient adaptation of pre-trained language models. However, existing methods typically rely on long continuous prefixes, which not only occupy the limited context window but also reduce effective input capacity. Furthermore, aggressively compressing prompt length introduces a severe capacity gap, leading to significant performance degradation. To address this challenge, we propose LR-HKD, a low-rank hybrid knowledge distillation framework for prompt compression. This framework adopts an alternating layer strategy, leveraging low-rank structural alignment to capture mini-batch-level inter-sample relationships, while employing intermediate representation matching to model fine-grained point-wise features. Extensive evaluations on multiple benchmark tasks demonstrate that LR-HKD can achieve substantial prompt compression while maintaining strong downstream task performance. Notably, the distilled student model not only outperforms existing baselines but, on certain tasks, even surpasses the uncompressed teacher model. These results highlight the effectiveness of combining feature-level and structure-level distillation for achieving robust knowledge transfer under extreme representational constraints.

Keywords: Transfer Learning, Deep Learning, Representation Learning
Abstract: Text-to-SQL is a fundamental task in natural language processing that enables users to interact with structured databases using natural language. While large language models (LLMs) have demonstrated remarkable performance on this task, their substantial computational requirements hinder deployment in resource-constrained settings. In this paper, we introduce textbf{SQuaD-SQL} (textit{Small-Qualified and Distilled for SQL}), a novel approach that empowers small language models (SLMs) to approach the performance of LLMs on the Text-to-SQL task while significantly improving efficiency through knowledge distillation and synthetic data generation. Our method comprises three key components: (1) textit{LLM-based synthetic data generation}, where structured knowledge is extracted from LLMs via carefully designed prompting strategies; (2) textit{parameter-efficient fine-tuning}, enabling full model training on a single consumer-grade GPU; and (3) textit{domain-adaptive fine-tuning}, where domain-specific synthetic data further enhances performance in targeted domains. Experiments on the WikiSQL dataset demonstrate that SQuaD-SQL achieve execution accuracies of 86.9% in Test data, respectively, approaching the performance of LLMs, while offering faster inference and lower memory usage. These results suggest that, with proper training strategies, SLMs can serve as practical and efficient alternatives for Text-to-SQL applications in resource-limited environments.

Keywords: Deep Learning, Image Processing and Pattern Recognition, Neural Networks and their Applications
Abstract: Constrained by the computational bottlenecks of agricultural edge devices, existing models struggle to balance high precision and ultra-low overhead for cotton boll segmentation in complex scenarios. To address this issue, this paper proposes DSM-YOLO, a novel instance segmentation network based on YOLO11n-seg. The model introduces a Dynamic Feature Decoupling Module (C3K2-GDW) to adaptively extract complex contours and integrates a Multi-Scale Residual Channel Attention (MSRC) to suppress noise while preserving weak target features. Furthermore, a Lightweight State Space Mask Head (Segment_LSSM) is innovatively designed to achieve efficient global decoding while maintaining linear complexity. Experimental results on a self-built dataset consisting of 3,000 images, featuring challenging scenarios such as dense stacking and occlusion, demonstrate that DSM-YOLO contains only 2.40M parameters and 8.6 GFLOPs, achieving a peak mask mAP50-95 of 0.753. Compared to the baseline YOLO11n-seg, DSM-YOLO significantly improves the recall and mAP50 by 4.3% and 4.9%, respectively, while maintaining robust precision despite reductions in both parameter count and computational cost. This substantially reduces missed detections in dense occlusion scenarios, achieving an optimal balance between high-precision segmentation and extreme lightweight design.

Keywords: Deep Learning, Transfer Learning, Image Processing and Pattern Recognition
Abstract: To address the limitations of conventional self-distillation methods, including the reliance on a single deep-layer supervision scheme and insufficient decoupling granularity, this paper proposes a Cross-Layer Generalized Decoupled Self-Distillation (CL-GDSD) framework. Specifically, we incorporate generalized decoupled knowledge distillation into the self-distillation paradigm and introduce a Top-k-based category grouping mechanism to model high-probability confusing classes and low-probability long-tail classes in a more fine-grained manner. Furthermore, a multi-teacher cross-layer distillation strategy is designed to enable shallow classifiers to progressively acquire knowledge from intermediate layers, thereby alleviating the semantic gap between shallow and deep representations. Experimental results on CIFAR-10 and CIFAR-100 demonstrate that the proposed method consistently outperforms baseline models and existing self-distillation approaches across various ResNet architectures, achieving stable improvements in both shallow branch performance and overall ensemble accuracy.

Keywords: Generative AI, Reinforcement Learning, AI and Applications
Abstract: In recent years, the construction of world models by harnessing the powerful sequence modeling capabilities of Transformers has emerged as a prominent research direction. Although Transformer-based agents (e.g., STORM) have demonstrated high data efficiency in discrete environments, their performance in continuous environments remains suboptimal. This limitation stems primarily from the inherent parallel training mechanism of Transformers, which often processes image encoding as an independent, context-agnostic step, thereby diminishing the capacity of the learned representation space. To address this issue, we propose the Recurrent Transformer State Space Model (RTSM), a novel serially-trained Transformer-based world model designed to restore context-aware hybrid deterministic and stochastic state representations. Furthermore, to enhance training efficiency, we incorporate a KV Cache mechanism during training. Experimental results on the DeepMind Control Vision benchmark show that our method achieves significant performance improvements compared to baselines, while maintaining acceptable computational costs.

Keywords: Autonomous Vehicle, Intelligent Transportation Systems, Robotic Systems
Abstract: A key challenge in off-road planning is converting terrain recognition into a continuous cost representation that is both physically meaningful and useful for decision-making. Existing approaches usually rely on discrete semantic labels, heuristic cost rules, or black-box learned scores, which do not adequately capture intra-class terrain variation and are difficult to interpret across different vehicles and environments. This paper proposes a physics-constrained continuous cost modeling framework based on a vision-modulated terrain parameter field. Semantic segmentation provides terrain-class priors, while visual cues are treated as observational proxies that refine terramechanics parameters under bounded and monotonic constraints. The estimated parameter field is then propagated through traction, slope, sinkage, and vibration models to compute safety margins, which are transformed into a unified risk-aware cost field and integrated into Hybrid A* as soft constraints. Experiments show that, compared with Hybrid A* without cost guidance, the proposed method reduces the average path cost by 33.3%, and reduces the exposure to hazardous-region exposure ratio by 61.7%; compared with Theta*, which is the strongest baseline in average cost, it still achieves a 24.8% reduction. In local perturbation experiments, replanned paths further reduce average path cost by 18.5%, demonstrating robust avoidance of newly emerged high-risk regions.

Keywords: Deep Learning, Machine Learning, Neural Networks and their Applications
Abstract: Cyberattacks continue to evolve, and Intrusion Detection Systems (IDS) must contend with high-dimensional network traffic alongside severe class imbalance in standard datasets such as NSL-KDD. To tackle these challenges, we introduce CBiLSTM-Att—a hybrid deep learning architecture that combines convolutional neural networks for localized feature extraction, bidirectional long short-term memory for sequential context, and a self-attention mechanism to dynamically focus on relevant features. Hyperparameter selection, including the number of LSTM units and convolutional filters, is handled through the Sparrow Search Algorithm, which balances optimization efficiency with model performance. Data preprocessing follows a stratified split, and class weighting is applied to improve detection of underrepresented attack types. Evaluated on NSL-KDD, the model attains 99% accuracy for binary classification and 92% for five-class classification, with recall rates of 87.50% for U2R attacks and 94.33% for R2L attacks. The main contribution is a reproducible framework that handles data imbalance through integrated optimization and attention-driven feature selection.

Keywords: Expert and Knowledge-Based Systems, Generative AI, Cognitive Computing
Abstract: Knowledge graph question answering (KGQA) systems uniformly apply expensive multi-hop traversal to every query regardless of complexity. We propose DualGraph, a dual-process framework inspired by Kahneman's fast/slow cognition that routes queries between two graph reasoning modes: a fast pattern-cache path (System 1) for familiar, structurally simple queries and a deliberative beam-search traversal (System 2) for complex ones. A classifier-free familiarity gate—requiring no gradient-based training beyond a single threshold tuned via grid search—combines template confidence, graph-structural confidence, and draft-answer entropy to make routing decisions using graph-side signals absent from syntax-only routers. Unlike prior adaptive methods that route between text and graph retrieval, both modes of DualGraph operate on the knowledge graph. On four KGQA benchmarks, DualGraph matches state-of-the-art prompt-based accuracy while consuming 53% fewer tokens and reducing average wall-clock latency by 41% (from 8.0s to 4.7s per query) compared to uniform multi-hop traversal.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Deep Learning
Abstract: Forward-looking sonar provides stable observations for underwater perception, but object detection is challenging due to speckle noise, low contrast, and weak target boundaries. Existing detection models often discard fragile contour cues and lack the ability to capture long-range dependencies in cluttered acoustic environments. This paper introduces HESonarNet, a contour-aware hybrid framework that unites shallow structural preservation with deep contextual reasoning. The proposed method utilizes an Edge-Guided Enhancement Module (EGEM) to extract and refine weak contour responses from shallow layers. Meanwhile, a Transformer encoder combined with a multi-scale perception bridge models global contextual information from deep features. These representations are integrated using a bidirectional cross-modal attention fusion module to align structural details with semantic context. Experiments conducted on the Marine debris dataset and UATD benchmarks validate the framework. HESonarNet achieves notable improvements in overall detection accuracy and spatial localization compared to existing detectors, successfully reducing false alarms and missed targets in complex marine scenes.

Keywords: Swarm Intelligence, Logistic and Manufacturing Automation, Evolutionary Computation
Abstract: The Capacitated Vehicle Routing Problem (CVRP) is a challenging combinatorial optimization task in logistics scheduling. To address the limitations of traditional heuristic algorithms, which often get trapped in local optima and exhibit slow convergence on large-scale instances, this paper proposes DGPACO, a deep graph prior-guided heterogeneous ant colony optimization framework. By integrating the high-dimensional feature extraction capability of neural networks with the strong exploitation capability of swarm intelligence, the proposed method is able to effectively optimize the search trajectory. First, an EdgeConv-based neural heuristic prior captures global topological structures, overcoming the myopic nature of static distance heuristics. Second, a prediction-guided pheromone warm-start mechanism enables the algorithm to bypass early blind exploration and directly target high-quality solution spaces. Finally, a heterogeneous multi-colony swarm strategy dynamically maintains population diversity while strengthening local exploitation, preventing premature convergence. Extensive experiments and ablation studies conducted on scalable synthetic datasets demonstrate that DGPACO achieves rapid convergence in minimal iterations. It strikes a good balance between computational overhead and solution accuracy, exhibiting superior robustness and global search capabilities in large-scale complex routing networks compared to state-of-the-art algorithms.

Keywords: System Modeling and Control, Autonomous Vehicle, Application of Artificial Intelligence
Abstract: Accurate vehicle dynamics modeling is essential for the development of advanced safety and autonomous driving systems. However, obtaining precise mathematical models for small-scale commercial vehicles is often unfeasible due to the lack of access to proprietary physical parameters and complex tire-road interactions. This paper proposes a novel data-driven system identification architecture: the Convolutional Neural Network with Output Recurrence and Exogenous Input (CNN-OREI). The proposed model leverages 1D-convolutional layers for automated feature extraction from time-series data, combined with an output recurrence mechanism to establish robust temporal dependencies. The network’s topology, specifically the number of convolutional filters and hidden neurons, was optimized using Particle Swarm Optimization (PSO). Experimental results demonstrate that the CNN-OREI outperforms traditional linear transfer functions, state-space representations, and standard NARX neural networks. The proposed architecture achieved a Fit of 95.99% and a Normalized Root Mean Squared Error (NRMSE) of 0.0117 on validation data, maintaining a stable prediction error within ± 0.01 m/s. These findings validate the CNN-OREI as a highly effective solution for identifying the intricate longitudinal dynamics of small-scale electric vehicles, providing a reliable foundation for real-time control and navigation strategies.

Keywords: AI and Applications, Medical Informatics
Abstract: Fetal growth restriction (FGR) is a leading cause of perinatal morbidity and fetal mortality, making early prediction of FGR clinically and socially valuable. The use of artificial intelligence techniques to analyze ultrasound imaging data represents an effective approach for FGR prediction. However, due to the difficulty in identifying FGR-relevant regions of interest in early ultrasound data, existing research has primarily covered FGR detection after 18 weeks of gestation, which is relatively delayed for clinical intervention. To address this issue, we propose a hybrid multi-stage feature learning-based FGR prediction method. First, a fine-tuned U-Net network is employed to extract highly FGR-relevant regions of interest from the ultrasound data. Subsequently, a dual-paradigm architecture is utilized to achieve complementary feature learning on these regions of interest. Finally, a weighted decision fusion module is designed to fuse the extracted features. Experimental results demonstrate that our method can accurately identify cases ultimately diagnosed with FGR from first-trimester (11-13 weeks) placental ultrasound images, achieving an AUC of 0.741. Meanwhile, on a placental ultrasound dataset acquired between 18-24 weeks, it achieves higher prediction accuracy than state-of-the-art models, with an AUC of 0.845.

Keywords: Biometrics and Applications, Generative AI, Medical Informatics
Abstract: Medical image segmentation is fundamental to clinical analysis but remains challenging due to complex backgrounds, ambiguous boundaries, and large morphological variations. Although diffusion models excel in generative tasks, their high sampling costs and training instability limit their application in segmentation. To address this issue, we propose FlowSeg, an efficient mask generation framework based on Conditional Flow Matching (CFM). Unlike diffusion based approaches, FlowSeg directly models continuous generative trajectories, significantly improving inference speed and training stability while maintaining high generation quality. Moreover, to address the inherent boundary ambiguity in medical images, a parallel shape constrained flow and a gated fusion module are designed to explicitly enhance the model’s capability to capture fine-grained boundary details. Extensive experiments on the MoNuSeg, ISIC, and a private dataset demonstrate that FlowSeg consistently surpasses state-of-the-art methods in accuracy and efficiency. This work provides a novel, reliable solution for efficient medical image segmentation with strong potential for clinical translation. The source code will be made publicly available upon publication.

Keywords: Optimization Control, Game Theory and Multi-Agent Systems, Decision Support Systems
Abstract: 奖励广告是平台经济中新兴的形式，服务定价和广告量显著影响用户的战略行为。本研究聚焦于在免费增值模式下运营的双头视频平台，基于奖励广告策略构建了一个博弈论模型。我们推导出四种策略组合中最优的增值服务（VAS）价格和广告量。数值分析进一步探讨了增值服务奖励水平如何影响最佳定价、广告量和最大收入，并在不同增值服务绩效水平下识别均衡策略。结果显示，奖励水平决定了最佳结果，且其效果在不同广告策略之间存在差异。当两个平台之间的增值服务质量差距较大时，竞争平台应采用对称的奖励广告策略。本研究为视频平台在优化增值服务定价和广告投放方面提供了理论框架和决策支持。

Keywords: Deep Learning, Neural Networks and their Applications, Intelligent Industrial Systems
Abstract: Industrial time series forecasting is pivotal for intelligent decision-making but remains fundamentally challenged by non-stationarity and continuous distribution drift. Existing transformer-based models often rely on passive static normalization, failing to adapt to evolving data distributions during inference. To address this limitation, we propose a novel framework for Proactive Distribution Adaptation, shifting the paradigm from passive statistical fitting to active distribution alignment. The framework integrates three key components: (1) an Affine Prototype-Timestamp (APT) module that retrieves temporal prototypes to perform dynamic affine transformation, stabilizing feature representations against instantaneous fluctuations; (2) a Future-Guided Learning (FGL) mechanism using feature-level distillation to align latent representations with future distribution trajectories; and (3) a Page-Hinkley (PH) monitor that detects real-time residual deviations to trigger adaptive model updates. Experiments on real-world industrial datasets demonstrate that our approach significantly outperforms state-of-the-art baselines in predictive accuracy and robustness. These results validate the efficacy of combining prototype-based alignment with future-guided distillation for robust forecasting under complex distribution shifts.

Keywords: Application of Artificial Intelligence, Artificial Social Intelligence, Deep Learning
Abstract: Monocular vision offers a low-cost route toward scalable cardiopulmonary resuscitation (CPR) training, but reliable quality assessment remains difficult because clinically relevant cues such as compression depth, recoil, and upper-body force transmission are only indirectly observable in RGB videos. This paper presents an assessment-oriented monocular CPR evaluation framework that couples task-adapted front-end perception with a clinically constrained scoring back-end. The front-end stabilizes downstream analysis through CPR-oriented instance segmentation and phase recognition, while the back-end reconstructs 3D skeletal motion with a topology-constrained GraphMLP-CPR model and evaluates quality through a deviation-aware dual-track scorer that combines explicit clinical descriptors with a deep skeletal branch. We do not position the front-end adaptations as universal new backbones; rather, the contribution lies in how these components are reconfigured around downstream assessment reliability. Experiments on a self-built CPR dataset with synchronized mannequin-based reference labels show that the improved scene parser raises FastInst from 82.8% to 84.5% AP while increasing speed from 53.8 to 62.4 FPS, the phase recognizer attains 94.0% mAP at 14.8 GFLOPs, GraphMLP-CPR reduces MPJPE from 45.5 to 44.8 mm relative to GraphMLP, and the final dual-track evaluator reaches 95.4% accuracy with a Spearman correlation of 0.947. These results suggest that clinically grounded coupling between perception and scoring is a practical design strategy for monocular CPR training systems.

Keywords: Deep Learning, Machine Learning, Computational Life Science
Abstract: The accurate identification of Drug-Target Inter actions (DTIs) is crucial for advancing drug discovery and repurposing. Although existing Graph Neural Network (GNN) based methods have shown promise in DTI prediction, they primarily focus on pairwise node interactions and struggle to capture deeper, higher-order dependencies. Furthermore, current hypergraph representation learning faces challenges in handling highly heterogeneous and complex relational pat terns. To address these limitations, we propose MMHC-DTI, a novel hypergraph-based framework for DTI prediction. The framework leverages motif-driven hypergraph construction to characterize complex high-order relationships among drugs. A Multi-scale Motif Hypergraph Convolution (MMHC) module is designed to simultaneously capture local structural rela tionships and long-range interaction dependencies through multi-hop propagation and a scale attention mechanism. To overcome the semantic isolation inherent in individual hyper graph branches, an Inter-branch Semantic Correlation Module (ISCM) is further introduced to facilitate information exchange across disparate semantic subspaces. Experimental results demonstrate that MMHC-DTI consistently outperforms state of-the-art baselines across all key evaluation metrics, validating the effectiveness and superiority of the proposed method in modeling complex DTIs.

Keywords: System Modeling and Control, Intelligent Industrial Systems
Abstract: Accurate system modeling is an essential prerequisite for the high-performance industrial servo control. In particular, permanent magnet synchronous motor driven servo system is recognized as a representative application, where parameter identification is critical. However, traditional simulation-driven identification methods commonly explore parameter regions prone to encounter simulation failures, resulting in numerical divergence, physical inconsistency, and low computational efficiency. To address such issues, this paper proposes a failure-robust constrained Bayesian optimization (FRCBO) framework for industrial servo system modeling. Specifically, FRCBO adopts a novel decoupled dual-surrogate architecture, utilizing a Gaussian process regressor to model the tracking error and a Gaussian process classifier to estimate the simulation feasibility. On top of this, we further design a feasibility-weighted expected improvement acquisition function to penalize risky parameter configurations, so as to guide the search away from divergent and physically inconsistent solutions. Both simulations and real-world experiments on an aerospace servo testbed are conducted, which demonstrate that the proposed method can significantly reduce simulation failures while achieving superior identification accuracy and data efficiency over existing methods.

Keywords: Affective Computing, Deep Learning, Machine Learning
Abstract: Abstract— Multimodal Emotion Recognition in Conversations (MERC) based on textual, acoustic, and visual modalities requires both multimodal understanding and conversational reasoning. However, existing decomposition methods often rely on coarse component partitioning, and current MERC models still lack effective mechanisms to organize heterogeneous emotional cues and capture high-order conversational dependencies. To address these limitations, we propose Decomposition-guided Structured Hypergraph Reasoning (DSHR), which decomposes multimodal cues into unique (U), pairwise (P), consensus (C), and synergistic (S) components. Based on these components, DSHR constructs consensus and synergistic hyperedges, while using modality-specific discrepancies to suppress unreliable connections. In this way, DSHR explicitly links representation decomposition with relational structure construction. Experiments on two benchmark datasets demonstrate the effectiveness, stability, and generalizability of DSHR. Compared with large-scale PLM/LLM-based methods, DSHR offers a lightweight and interpretable alternative without relying on expensive instruction-tuned language models.

Keywords: Deep Learning, Generative AI, Biometric Systems and Bioinformatics
Abstract: A primary bottleneck in discrete molecular graph diffusion is maintaining topological coherence. Standard homogeneous noise schedules treat all graph elements equally, which often causes severe global structure degradation (e.g., ring cleavage) at high noise levels. In response, we present Hierarchy-Aware Diffusion for Topology-Preserving Molecular Design (Motif), a framework that leverages the Bemis–Murcko scaffold hierarchy as a guiding structural prior. Central to our approach is a Bipartite Noise schedule that distinctly regulates the diffusion dynamics for scaffolds versus decorators. By preserving a higher signal-to-noise ratio for core scaffolds, Motif guarantees a coarse-to-fine generation paradigm, anchoring the global topology prior to local refinement. Additionally, we introduce Topological Constraint Reinforcement (TCR), a customized reconstruction objective designed to counteract sparsity-induced training imbalances by selectively routing gradients to critical topological substructures. Empirical results across standard multi-conditional polymer and small-molecule benchmarks confirm that Motif achieves superior Fréchet ChemNet Distance (FCD) and enhanced property control compared to strong diffusion baselines. We conclude with a comprehensive evaluation of the balance between global structural resilience and local detail accuracy.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Multimedia Computation
Abstract: Efficiently retrieving specific 3D instances from large-scale scenes via natural language prompts remains a formidable challenge in multimedia analysis. Existing approaches predominantly adhere to a ``scene-level embedding'' paradigm, which necessitates distilling high-dimensional semantic features into every 3D primitive. This strategy suffers from a fundamental architectural bottleneck: the memory and computational costs scale linearly with scene complexity, inevitably triggering Out-Of-Memory (OOM) failures in city-scale environments. To dismantle this barrier, we propose QueryGaussian, a training-free framework enabling expeditious and scalable open-vocabulary 3D instance retrieval. Diverging from holistic semantic distillation, QueryGaussian employs a novel Instance-Level Query Mechanism that rigorously decouples semantic understanding from geometric representation. Specifically, we leverage pre-trained 2D vision models to interpret user prompts and lift segmentation masks into 3D via a concurrent maximum-weight association strategy, ensuring strict semantic-visual consistency. To mitigate projection ambiguity, we introduce a temporal fusion module with multi-stage adaptive density clustering. Experimental results demonstrate that QueryGaussian not only matches the accuracy of state-of-the-art methods but also delivers a decisive efficiency leap—reducing GPU memory usage by over 70% and accelerating inference by 180x. Crucially, QueryGaussian enables expeditious instance retrieval on city-scale scenes (containing tens of millions of Gaussians) on consumer-grade hardware.

Keywords: Environmental Sensing, Robotic Systems, Reinforcement Learning
Abstract: To address the challenges of humanoid robots performing inclined ladder grasping tasks in complex and dynamic environments, this paper proposes a Spatio-Temporal Attention-based Multimodal Fusion Deep Reinforcement Learning framework (STFusion-Net). Specifically, a channel attention mechanism is introduced in the visual encoding stage to enhance critical ladder-step features. A spatiotemporal attention fusion module is further designed to model cross-modal interactions among visual observations, joint angles, and IMU signals. The fused representations are then integrated into a Dueling Deep Q-Network to enable stable policy learning. The proposed method is evaluated in the Webots simulation environment against unimodal baselines (vision-only and joint-only) as well as a simple concatenation-based fusion approach. Experimental results demonstrate that STFusion-Net achieves the highest grasping success rates in both static and dynamic scenarios (99% and 90%, respectively), while exhibiting more stable performance. These results indicate that STFusion-Net can effectively leverage multimodal information, providing robust policy guidance for high-precision motion execution and enabling reliable perception and decision-making for humanoid robots in complex environments.

Keywords: Human-Machine Interaction, Image Processing and Pattern Recognition, Multimedia Computation
Abstract: Vision-Language Tracking aims to locate a specified target within a video sequence based on a natural language description. However, the performance and generalization of existing methods are limited by the lack of linguistic diversity in training data, especially when handling implicit, reasoning-based instructions that are common in real-world scenarios. To address this, we extend existing benchmarks with reasoning-oriented descriptions, resulting in a reasoning corpus for training and the textbf{ReasonTrack-Bench} test set of 882 manually verified sequences for evaluation. We also propose textbf{R2Track}, a underline{textbf{R}}easoning-aware and underline{textbf{R}}obust underline{textbf{T}}racking framework integrating a Multimodal Large Language Model (MLLM) adapter. This adapter functions as a language interpreter, translating complex queries into simplified cues for a downstream tracker. The adapter is trained via a two-stage strategy of Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) to enhance localization accuracy and language understanding. Experiments on ReasonTrack-Bench show R2Track significantly outperforms existing methods, demonstrating superior generalization to diverse and complex language.

Keywords: Intelligent Green Production Systems, Game Theory and Multi-Agent Systems, Service Systems and Organizations
Abstract: ：本文探讨了不同内容的影响 权力结构对决策和社会福利的影响 考虑政府建设双渠道绿色供应链 研发补贴。通过表述 由制造商主导和零售商主导的Stackelberg游戏模型， 我们推导出两者均下的均衡结果 价格一致和价格不一致的情景。该 研究结果显示，权力结构显著 影响补贴的有效性。在 价格一致的情景，制造商主导的结构 更有利于提升整体供应链利润， 产品绿色环保和社会福利。然而，在 价格不一致情景，零售商主导结构 表现出优异的性能，从而获得更优的值 所有指标均与价格一致情景进行比较。这项研究为政府提供了理论支持 在设计绿色补贴政策时，并提供重要服务 供应链企业的管理洞察 在不同权力结构下的决策。