Depression is a serious mental health disorder with a potential hazard for individuals and society characterized by persistent sadness and hopelessness. Multimodal information including vision, audio, and text is critical for depression diagnosis and treatment. Most studies focus on designing sophisticated feature extraction methods but ignore feature enhancement and fusion within intra-modality and cross-modality. In this paper, a Chinese Multimodal Depression Corpus (CMD-Corpus) dataset is established assisted by clinical experts aiming to support more depression research. Furthermore, we propose a multimodal depression recognition framework based on Mutual Information Maximization with Multi-task Learning (MIMML) to enhance feature representation and fusion among video, audio, and text modalities. The MIMML employs the strategy of maximizing mutual information to accelerate modality-invariance enhancement. The multi-task is used to improve the representation performance of the single modality to improve modality-specific enhancement. Meanwhile, a gated structure with bidirectional gated recurrent units and convolutional neural networks is designed to achieve multimodal feature fusion, which is key to boosting completeness among modalities. Experimental results show that the proposed MIMML effectively captures representation to increase depression recognition accuracy, achieving 84% and 89% accuracy on DAIC-WOZ and our self-collected CMD-Corpus dataset respectively.

Multi-Provider Edge Cloud (MPEC) is an important 5G network paradigm where Virtual Network Functions (VNFs) can be deployed on capacity-limited edge servers from different providers and near users, delivering low latency services. Compound Service Function Chains (SFCs) are ordered chains consisting of user-requested VNFs and Supplementary VNFs (S-VNFs) that bridge requested VNFs from different providers. However, different deployment schemes of requested VNFs determine the presence of distinct S-VNFs with varying capacity demands and processing delays. How to deploy VNFs of compound SFCs on edge servers and route their traffic through these VNFs within MPEC networks to minimize total service delay poses a great challenge. In this paper, we first formulate the compound SFC deployment problem as an integer nonlinear program and prove its NP-hard. To solve this problem, we analyze the components of service delay, decompose the problem into three subproblems and propose a heuristic algorithm. We further consider a condition that collaborations among providers are constrained to a limited scope and devise a reinforcement learning-based scheme to address this condition. Next, we present a theoretical analysis of the two proposed algorithms, including time complexity and worst-case performance bounds. Finally, extensive experiments on real-world datasets demonstrate that our algorithms outperform several existing methods.

Automatic sleep staging plays a vital role in assessing sleep quality and diagnosing sleep disorders. Most existing methods rely heavily on long and continuous EEG recordings, which poses significant challenges for data acquisition in resource-constrained systems, such as wearable or home-based monitoring systems. In this paper, we propose the task of resource-efficient sleep staging, which aims to reduce the amount of signal collected per sleep epoch while maintaining reliable classification performance. To solve this task, we adopt the masking and prompt learning strategy and propose a novel framework called Mask-Aware Sleep Staging (MASS). Specifically, we design a multi-level masking strategy to promote effective feature modeling under partial and irregular observations. To mitigate the loss of contextual information introduced by masking, we further propose a hierarchical prompt learning mechanism that aggregates unmasked data into a global prompt, serving as a semantic anchor for guiding both patch-level and epoch-level feature modeling. MASS is evalutaed on four datasets, demonstrating state-of-the-art performance, especially when the amount of data is very limited. This result highlights its potential for efficient and scalable deployment in real-world low-resource sleep monitoring environments.

Although multi-agent reinforcement learning (MARL) has achieved significant success in various domains, its deployment in real-world scenarios remains challenging, particularly in communication-constrained environments involving multi-task coupling. Existing methods suffer from two limitations: (1) the inability to effectively integrate and process incomplete state from disparate agents, and (2) a lack of robust mechanisms for handling complex multi-task coupling. To address these challenges, we propose the Coupled Communication-Task Decoupling (CCTD) framework. CCTD introduces two critical innovations: first, a distributed state compensation mechanism to process historical data, thereby reconstructing accurate global states from partial observations; second, a hierarchical architecture that systematically decomposes complex tasks into manageable subtasks while preserving their interdependencies. Thanks to its modular design, CCTD can integrate with existing MARL algorithms and allow for flexible combination of various subtasks. Extensive experiments demonstrate that CCTD outperforms baseline methods, achieving a 10% improvement in communication reception rate and superior performance across all subtasks in multi-task environments.

This paper investigates digital twin-aware task offloading in vehicular edge–cloud networks with stochastic arrivals and time-varying wireless channels, with the objective of minimizing long-term power consumption while ensuring delay-aware queue stability. We formulate an online joint control of the association, communication, and computation, and further propose a Lyapunov-guided deep reinforcement learning (LyDRL) algorithm. The algorithm adopts a two-stage controller, where a policy-based actor determines vehicle–server association, while a model-driven critic performs lightweight analytical updates of uplink bandwidth, offloaded tasks, and processing frequencies across the vehicle, edge, and cloud. These design choices result in three advantages: (i) virtual quality-of-service (QoS) queues enforce queue stability; (ii) a drift-plus-penalty weight makes the power–latency trade-off explicit and tunable; and (iii) digital twin prediction errors and channel uncertainty caused by imperfect channel state information (CSI) are embedded in the control state to improve robustness in online decision-making. Consequently, the decision routine runs in linear time per slot with modest memory overhead, enabling real-time deployment at scale over heterogeneous infrastructures. Trace-driven simulations on a $5\times 5$ Manhattan grid demonstrate that LyDRL achieves rapid convergence and consistently outperforms purely data-driven DRL baselines. Specifically, it tightly bounds the physical and virtual queue backlogs at low levels under heavy traffic loads, minimizes system power consumption, and exhibits strong resilience against both digital twin prediction errors and CSI estimation errors.

Preference optimization is widely used to align large language models (LLMs) with human preferences, yet many margin-based objectives often suppress the chosen response together with the rejected one, and no general mechanism exists to prevent this across objectives. We bridge this gap by presenting a unified \textbf{incentive-score decomposition} of preference optimization, revealing that diverse objectives share identical local update directions and differ only in their scalar weighting coefficients. Building on this decomposition, by analyzing the dynamics of the rewards of chosen/rejected responses, we identify the \textbf{disentanglement band (DB)}, a simple, testable condition that characterizes when training can realize the ideal pathway: suppressing the loser while maintaining the winner, possibly after an initial transient. Leveraging the DB, we propose a plug-and-play \textbf{reward calibration (RC)} that adaptively rebalances chosen versus rejected updates to satisfy the DB, without redesigning the base objective. Empirical results confirm that this calibration effectively disentangles updates and improves alignment performance across diverse objectives.

Chemical-Disease-Gene (CDG) association prediction-encompassing Chemical-Disease (CD), Disease-Gene (DG), and Chemical-Gene (CG) interactions-is a cornerstone of drug discovery, as it underpins target identification and drug repurposing. While these tasks are inherently synergistic, existing methods often address them in isolation, failing to capture shared heterogeneous semantics and cross-task dependencies. We hypothesize that a unified pretraining framework can learn transferable biomedical semantics across CDG tasks, with task-specific prompt tuning enabling efficient adaptive fine-tuning without full retraining. To test this hypothesis, we propose the Pretraining-Prompt-Finetuning Heterogeneous Graph Neural Network (PPF-HGNN), a two-stage framework built on heterogeneous graph neural networks (GNNs) and prompt learning. Specifically, we construct a CDG heterogeneous graph, employ parameter-free metapath-guided message passing for high-order semantic capture, and optimize generalizable representations via a dual self-supervised objective (association prediction + feature reconstruction) during pretraining. For downstream tasks, task-specific learnable prompt vectors are introduced to adapt frozen pretrained representations to CD, CG and DG association prediction tasks via additive fusion, preserving core semantics while injecting task-specific biases. Comprehensive experiments demonstrate PPF-HGNN's state-of-the-art performance: AUC of 0.9633 (CD), 0.9939 (CG), and 0.9390 (DG), with F1-scores of 0.9157, 0.9668, and 0.8955 respectively-substantially outperforming six existing baselines. This work validates the pretrain-prompt-finetune paradigm for multi-task biomedical association prediction, providing a robust AI-driven tool to accelerate translational research and decipher complex CDG relationships. The source code is available at https://github.com/ike-zengxi/PPF-HGNN.

The advancement of emerging technologies such as uncrewed aerial vehicles and distributed sensing devices has accelerated the development of Autonomous Intelligent Systems (AIS) with integrated perception, communication, and computation capabilities. However, in mission-critical consumer applications such as smart cities, a major challenge lies in designing a computational architecture that combines deployment flexibility with domain-specific expertise, enabling continuous self-evolution and collaborative decision-making based on dynamic sensing data. To address this challenge, we propose a novel hierarchical cognitive architecture. The framework adopts a project-based methodology, where project engineers can autonomously construct systems by selecting utilities from a utility panel that integrates physical models, data-driven models, and AI models. The system's intelligence is powered by federated knowledge engine, which coordinates a set of domain-specific fine-tuned large language models (LLMs), and orchestrates their collaboration through a multi-agent workflow to execute complex tasks. We conduct extensive experiments demonstrating that, compared with several general-purpose LLMs, our fine-tuned model under our approach exhibits a significantly lower frequency of moderate-to-severe hallucinations in the water resources engineering domain. These findings validate the feasibility of constructing a federated knowledge engine. Furthermore, this work provides a robust and scalable platform for developing next-generation consumer-grade AIS that require high reliability and physical interpretability.

Transformer-based models have recently adopted increasingly complex structure (e.g., deeper or wider stacked network) to promote the representation learning capabilities of vision recognition. However, progressively deeper or wider stacked network cause the expensive computation cost, which hinders their effective deployment in resource-constrained edge clouds or end devices. In this paper, we propose DTSNet, a dynamic transformer slimming model, which scales vision transformers (ViTs) down across layers from both of the model depth and input width. This is the first time to explore the joint reduction of input tokens and model parameters for ViTs under maintaining performance. Specifically, DTSNet adopts a diversity-enhanced weight sharing module to reduce network parameters, where the weight knowledge of multiple adjacent blocks is effectively integrated into one block. Furthermore, DTSNet designs a unified and massively scalable token pruning mechanism that dynamically discarding less important tokens with a model-driven manner, by introducing a series of discriminant parameters, which is a simple change to the common architecture of vision transformers. Extensive experiments are conducted to verify that DTSNet is able to yield high efficacy in compressing parameter space and accelerating model inference. DTSNet-T/-S/-B on ImageNet achieves 3.0 M/11.1 M/42.9 M parameters and 0.8/2.9/13.7 GFLOPs, where number of parameters are reduced by 48% similar to 51% and inference speed are improved by 1.3 & times; similar to 1.5 & times;. Experiments results on semantic segmentation and object detection dataset further demonstrate the potential of DTSNet on complex dense prediction tasks

Human activity prediction (HAP) is crucial for enabling intelligent smart home services; yet, it is often hindered by the scarcity of high-quality, multidimensional datasets. Existing datasets are typically fragmented, capturing either long-term activity sequences or short-term device interactions, but rarely both in a unified manner. Traditional data collection methods are costly and time-consuming, while conventional simulation techniques struggle to generate diverse and logically coherent behavior sequences. To address these limitations, we propose SmartLLM, a novel large language model (LLM)-based simulation framework for automated generation of multidimensional smart home datasets. SmartLLM simulates simulated agents with distinct profiles (e.g., old man, remote worker, and holiday maker) performing daily activities within configurable home environments, generating temporally aligned sequences across activity-device-sensor dimensions. We generate two months of simulated data for three user profiles and validated their plausibility through activity distribution visualization, statistical perplexity analysis, and case studies. Multidimensional feature validation experiments further demonstrate that our multidimensional data significantly enhances the accuracy of activity prediction models compared to using single-dimensional features. This work successfully addresses key bottlenecks in smart home data acquisition and provides a scalable, high-quality data foundation for advancing smart home algorithm research.

The proliferation of hate speech on social media presents an escalating threat to both public discourse and individual mental well-being. Traditional strategies that prioritize detection and removal often neglect to engage directly with hate speakers or address the underlying causes of their hostility. This article proposes HateMediator, a dialogue-based intervention framework that fine-tunes large language models (LLMs) to generate persuasive, context-aware counter-hate speech. The framework emphasizes two core aspects: the generation of effective counter-hate responses and their evaluation through multiturn dialogues. Our fine-tuning approach integrates tutorial-based learning with critical token guidance, enabling LLMs to recognize and reproduce strategic rhetorical patterns observed in expert interventions. To support training and evaluation, we introduce the MedHate dataset, grounded in social science theory, comprising complete dialogue records from 85 real-world hate incidents (including 255 dialogues), expert-crafted counter-responses, and feedback from the original hate speakers. Experimental results show that HateMediator consistently outperforms baseline LLMs across multiple evaluation dimensions. This study advances both the technical frontier of hate speech intervention and the ethical deployment of LLMs in addressing complex social issues.

Reinforcement learning with verifiable rewards (RLVR) commonly uses entropy for advantage shaping. However, entropy cannot distinguish useful uncertainty from detrimental confusion, limiting its effectiveness as a correctness signal. We propose Contrastive Policy Optimization (CPO), which uses token-level contrastive disagreement between reference-guided and vanilla generation distributions for correctness-aware advantage shaping. Both theoretical and empirical results show that this disagreement reliably indicates token-level correctness. We further show that On-policy Distillation is a special case of CPO, where the posterior distribution is instantiated by an external teacher model. CPO also resolves the zero-advantage problem. Experiments on in-domain and out-of-domain benchmarks demonstrate that CPO substantially outperforms entropy-based RLVR methods while maintaining strong generalization. Further analysis shows that correct and incorrect responses naturally support exploration and exploitation respectively, and balancing both leads to the best performance.

With the advent of sixth-generation (6G) mobile communication technology, vehicle-to-everything (V2X) communication faces unprecedented challenges in communication efficiency, system generalization capabilities, and model collaboration. Conventional machine learning struggles with high-dimensional state spaces, slow convergence, and poor generalization under heterogeneous V2X nodes, rapidly varying channels, and multimodal sensing data in V2X systems. To address these issues, we propose a quantum-enhanced framework for V2X communication and model aggregation that targets efficient, robust, and intelligent transportation in 6G, which includes four modules: the channel-adaptive semantic communication module, the multimodal fusion module, the model transfer module, and the federated aggregation module. Specifically, the channel-adaptive semantic communication module leverages quantum convolutional neural networks (CNNs) and quantum distortion metrics to enable efficient transmission and strong generalization across diverse conditions. The multimodal fusion module exploits quantum attention and entanglement to compress features and associate semantics across heterogeneous data. The model transfer module employs quantum reinforcement learning to model decision making and improve adaptability in dynamic environments. The federated aggregation module integrates quantum tensor decomposition with backpropagation-based corrections to provide privacy preservation with low overhead and to strengthen global model robustness. This work outlines a new paradigm for communication and model collaboration in future 6G intelligent transportation.

Transformers are widely utilized in the point cloud domain. However, existing methods tend to overburden Transformer with the dual task of local geometric perception and global feature extraction, limiting its ability to capture high-level semantic knowledge. To address this issue, we present Representation Decoder (R-Decoder), a novel representation extraction module compatible with various point cloud Transformer methods, enabling the Transformer to focus on its excellent local perception. The R-Decoder iteratively extracts multiple global features from tokens generated by Transformer, refining them to construct an overall representation of point cloud. To ensure full adaptation of the R-Decoder to the knowledge of pre-trained Transformers, we design a cross-modal representation alignment task that leverages multimodal knowledge to specifically pre-train the R-Decoder. As a post-processing module, the R-Decoder seamlessly integrates with Transformers, while decoupling local perception and global representation. This design allows the Transformer to focus on the semantic encoding role for point tokens. Extensive experiments show that our R-Decoder significantly boosts the capabilities of 3D representation learning in various point cloud Transformer methods. Notably, it achieves impressive classification accuracies of 95.1% on the ScanObjectNN dataset and 95.3% on the ModelNet40 dataset. Moreover, our method obtains new SOTA on all benchmarks of few-shot and zero-shot classification, while enhancing the multimodal task capabilities of pre-trained Transformers.

In the context of quantum secure scenarios, existing research on mobile edge devices and intelligent computing and edge (ICE) systems based on the Non-Orthogonal Multiple Access (NOMA) communication model have overlooked the energy consumption overhead of Post-Quantum Cryptography (PQC) modules, and the high complexity of traditional resource allocation algorithms fails to meet the demands of real-time decision-making. To address these challenges, this paper proposes a lightweight agentic AI framework designed for online joint optimization within ICE-enabled mobile devices. The scheme constructs a multi-stage stochastic Mixed Integer Nonlinear Programming (MINLP) model that incorporates static power-consumption constraints for PQC modules. Based on Lyapunov optimization theory, the long-term optimization problem is decoupled, and a linear complexity algorithm is proposed to solve the nonconvex challenges of NOMA power allocation . Simulation results verify that the proposed scheme significantly improves computational throughput while ensuring system queue stability and energy consumption constraints. Compared with traditional Successive Convex Approximation (SCA) algorithms, the complexity is reduced to 𝒪(N), achieving a speedup of approximately 46 times when the number of devices N=35, thereby meeting the real-time decision-making requirements in dynamic wireless environments.

Visual emotion inference aims to predict human-perceived affective states from images, bridging low-level visual features with high-level emotional semantics. However, existing methods often rely on direct mappings from global features to emotion labels, lacking a mechanism to reflect the cognitive process by which humans perceive and interpret emotions. To address this limitation, we propose a dual-process visual emotion reasoning framework that integrates Multi-Patch Attention Flow (MPAF) and Fuzzy Emotional Semantic Modeling (FESM). MPAF is designed as the perceptual pathway of the proposed emotion reasoning framework, simulating how humans distribute attention and integrate emotional cues across patches. The image is divided into overlapping patches through which an attention flow dynamically integrates emotional cues, approximating how human vision distributes and refines attention across a scene. Meanwhile, FESM serves as the cognitive pathway of our frame-work based on fuzzy mathematics, projecting each patch into a continuous affective semantic space. Instead of fixed emotion labels, it assigns graded membership values across multiple fuzzy emotion concepts, formally simulating the ambiguity, continuity, and polysemy of human visual emotion reasoning. Within this unified framework, Patches of Interest (PoIs) naturally emerge as regions most aligned with the global affective state, providing interpretable insight into how local cues anchor overall emotional appraisal. Experimental results demonstrate that the proposed framework not only achieves effective emotion prediction but also offers cognitively meaningful interpretability, providing a promising step toward cognitively inspired visual emotion reasoning.

Magnitude-based stability proxies such as parameter drift are widely used in narrow-task fine-tuning, yet they do not reliably indicate degradation of broad capabilities. We identify trajectory lock-in: under fixed training conditions for narrow adaptation, the joint evolution of task loss and broad generalization collapses onto a shared low-dimensional degradation curve, so many stabilizers primarily change the rate of progress along this curve rather than altering the curve itself. This yields a drift paradox, in which comparable Euclidean displacement can still correspond to divergent generalization outcomes. To diagnose the underlying structure, we introduce objective-agnostic geometric probes that track the effective update subspace, together with an online harm signal that reflects curvature-dominated channeling toward directions associated with broad degradation. Finally, we show that escaping lock-in requires a spectral bifurcation, namely a qualitative reorientation of the update subspace toward softer curvature modes, thereby improving broad generalization while maintaining matched training performance. We validate these findings across model scales and modalities in narrow-task settings, and report practical deployment procedures and overhead measurements.

Preference optimization is widely used to align large language models (LLMs) with human preferences. However, many margin-based methods also suppress the chosen response when they try to suppress the rejected one, and there is no general way to prevent this across different objectives. We address this issue with a unified incentive-score decomposition of preference optimization, revealing that different objectives share the same local update directions and differ only in their scalar weights. This decomposition provides a common framework for analyzing objectives that were previously studied in separate settings. Building on this decomposition, by analyzing the dynamics of the chosen/rejected likelihoods, we identify the disentanglement band (DB), a simple, testable condition that tells us when training can follow the desired path: suppress the loser while preserving the winner, possibly after an early stage. Using the DB, we propose reward calibration (RC), a plug-and-play method that adaptively rebalances the updates for chosen and rejected responses to satisfy the DB, without redesigning the base objective. Empirical results show that RC leads to more disentangled dynamics, with better downstream performance observed across several settings. Our code is available at https://github.com/IceyWuu/DisentangledPreferenceOptimization.

The personalized fine-tuning of large language models (LLMs) on edge devices is severely constrained by limited computation resources. Although split federated learning alleviates on-device burdens, its effectiveness diminishes in few-shot reasoning scenarios due to the low data efficiency of conventional supervised fine-tuning, which leads to excessive communication overhead. To address this, we propose Language-Empowered Split Fine-Tuning (LESFT), a framework that integrates split architectures with a contrastive-inspired fine-tuning paradigm. LESFT simultaneously learns from multiple logically equivalent but linguistically diverse reasoning chains, providing richer supervisory signals and improving data efficiency. This process-oriented training allows more effective reasoning adaptation with fewer samples. Extensive experiments demonstrate that LESFT consistently outperforms strong baselines such as SplitLoRA in task accuracy. LESFT consistently outperforms strong baselines on GSM8K, CommonsenseQA, and AQUA_RAT, with the largest gains observed on Qwen2.5-3B. These results indicate that LESFT can effectively adapt large language models for reasoning tasks under the computational and communication constraints of edge environments.

The advanced Internet of Vehicles (IoV) integrates autonomous driving, intelligent navigation, and in-vehicle consumer electronics, relying on efficient and secure communications to ensure reliable services. While Reconfigurable Intelligent Surfaces (RIS) offer enhanced support for high-rate vehicular communication, existing RIS-aided scheduling schemes remain inefficient in delay-sensitive multi-task IoV scenarios and lack resilience against emerging AI-driven threats in consumer electronic systems. To overcome these limitations, this paper proposes a secure and communication-enhanced task offloading and resource allocation scheme for vehicular consumer electronic networks. First, we introduce an intelligent RIS-assisted vehicle model that jointly optimizes task dependencies, resource allocation, and RIS configuration under adversarial disturbances to improve transmission rate and communication quality. Second, to efficiently solve the resulting high-dimensional and non-convex optimization problem involving RIS phase shifts, we develop a Quantum-based Reinforcement Learning (QRL) framework. QRL encodes state spaces with quantum bits and embeds quantum circuits into neural networks, accelerating convergence and strengthening robustness against adversarial attacks. Experimental results show that QRL outperforms the Soft Actor-Critic (SAC) method in jointly optimizing offloading and allocation under security constraints. The proposed framework demonstrates strong generalization potential for various RIS-assisted mobile computing applications, especially in environments vulnerable to AI-powered threats.