Chenyang Wang

Lyapunov optimization plays a pivotal role in stabilizing systems and reducing energy consumption for online collaborative computing offloading. However, existing studies have encountered the issue of instantaneous greediness during the transformation process based on Lyapunov optimization, primarily due to the continuous operations involved in queue length calculations and local optimizations. This study investigates the application of soft actor-critic in optimizing the objective function of Lyapunov-guided queue models, aiming to achieve long-term and globally optimal offloading decisions. Firstly, the Lyapunov optimization framework is employed to build the service queue model and derive the upper bound of the system objective function. Secondly, the system action of resource allocation and the state space are constructed based on the queue model. Moreover, a deep reinforcement learning problem involving queue change and a Lyapunov drift-plus-penalty is formulated to minimize the system energy consumption. Thirdly, a queue-based internal discount factor is integrated into the system reward function of the SAC algorithm, enabling the system to achieve long-term gains more efficiently. Finally, an adaptive cloud-edge service offloading strategy is proposed to ensure queue stability and minimize energy consumption without necessitating a locally optimal solution at each time slot. Experimental results demonstrate that the proposed strategy exhibits superior scalability and robustness compared to other state-of the-art and baseline approaches in online collaborative service offloading systems.

In bandwidth-limited image transmission, maintaining perceptual quality at very low bit rates remains challenging. Conventional compression codecs typically introduce noticeable blocking artifacts and discard fine details when the compression ratio becomes too aggressive. To address this problem, we present BCP-SDXL, an edge-cloud collaborative image compression framework with cloud-side generative reconstruction. Instead of transmitting full-resolution pixels, BCP-SDXL sends two lightweight cues: a compressed Base image and a Canny structural map. On the edge side, an Adaptive Compression Agent, built on a Deep Q-Network (DQN), dynamically selects compression parameters and extracts structural priors with low deployment overhead. Upon reception, the cloud server leverages the Multimodal Large Language Model (Qwen3-VL-2B) to analyze the compressed base image and generate high-fidelity semantic descriptions. These descriptions, together with the base image and edge maps, are then used to condition a Latent Diffusion Model (SDXL) integrated with ControlNet. The resulting multi-condition reconstruction process restores missing high-frequency details while preserving the original scene’s structural layout. Experiments on public datasets show that BCP-SDXL reduces transmission cost by more than 90% and achieves better perceptual quality than conventional codecs and recent deep learning baselines under highly compressed settings.

Computing Power Networks (CPNs) represent an innovative, collaborative architecture that integrates resources via the communication network, optimizing resource allocation to support service demands. Due to the increased need for services powered by artificial intelligence across various domains, CPNs are increasingly required to allocate users efficiently to appropriate servers to meet the low-latency needs of service computing. However, challenges such as users' dynamic mobility, weak communication paths, and high-dimensional solution spaces persist in optimizing user allocation in CPNs. In this context, we propose a diffusion-driven optimization approach for mobility-aware user allocation. To tackle the challenge of users' dynamic mobility, we adopt a user location prediction approach incorporating the users' movement patterns to forecast future movement, called CAMPE. To tackle the challenge of weak communication paths, we establish the new transmission path by reconfigurable intelligent surface and enhance the quality of the communication link by adjusting the phase configurations. Moreover, faced with the challenge of high-dimensional solution spaces associated with phase adjustment and user allocation decisions, we devise an action-generation strategy based on diffusion models named DiffUser. This approach motivates the generation of optimal solutions even in complex and dynamic environments. Finally, we conduct extensive simulations in user location prediction and system latency optimization. Compared with other solutions, the superiority of our approach has been demonstrated.

Diffusion strategies have advanced visual motor control by progressively denoising high-dimensional action sequences, providing a promising method for robot manipulation. However, as task complexity increases, the success rate of existing baseline models decreases considerably. Analysis indicates that current diffusion strategies are confronted with two limitations. First, these strategies only rely on short-term observations as conditions. Second, the training objective remains limited to a single denoising loss, which leads to error accumulation and causes grasping deviations. To address these limitations, this paper proposes Foresight-Conditioned Diffusion (ForeDiffusion), by injecting the predicted future view representation into the diffusion process. As a result, the policy is guided to be forward-looking, enabling it to correct trajectory deviations. Following this design, ForeDiffusion employs a dual loss mechanism, combining the traditional denoising loss and the consistency loss of future observations, to achieve the unified optimization. Extensive evaluation on the Adroit suite and the MetaWorld benchmark demonstrates that ForeDiffusion achieves an average success rate of 80% for the overall task, significantly outperforming the existing mainstream diffusion methods by 23% in complex tasks, while maintaining more stable performance across the entire tasks.

To meet the demands of high-capacity and low-delay services, Fifth Generation (5G) Base Stations (BSs) are typically deployed in ultra-dense configurations, especially in urban areas. While this densification enhances coverage and service quality, it also leads to substantially increased energy consumption. However, the dense deployment pattern makes BS workloads more responsive to the spatiotemporal variations in user behavior, offering opportunities for energy-saving strategies that dynamically adjust BS operation states. In this context, we propose a Delay-aware and Energy-efficient Integrated Optimization System (DEIS) based on Deep Reinforcement Learning (DRL), which jointly optimizes energy consumption and network delay while maintaining user satisfaction. DEIS leverages a real-world dataset collected from operational 5G BSs provided by partner network operators, containing both BS deployment data and high-volume user request logs. Extensive simulations demonstrate that DEIS can achieve a 41% reduction in energy consumption while ensuring reliable delay performance.

Owing to its rapid progress and broad application prospects, few-shot action recognition has attracted considerable interest. However, current methods are predominantly based on limited single-modal data, which does not fully exploit the potential of multimodal information. This paper presents a novel framework that actively identifies reliable modalities for each sample using task-specific contextual cues, thus significantly improving recognition performance. Our framework integrates an Active Sample Inference (ASI) module, which utilizes active inference to predict reliable modalities based on posterior distributions and subsequently organizes them accordingly. Unlike reinforcement learning, active inference replaces rewards with evidence-based preferences, making more stable predictions. Additionally, we introduce an active mutual distillation module that enhances the representation learning of less reliable modalities by transferring knowledge from more reliable ones. Adaptive multimodal inference is employed during the meta-test to assign higher weights to reliable modalities. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing approaches.

Intent detection aims to identify user intents from natural language inputs, where supervised methods rely heavily on labeled in-domain (IND) data and struggle with out-of-domain (OOD) intents, limiting their practical applicability. Generalized Intent Discovery (GID) addresses this by leveraging unlabeled OOD data to discover new intents without additional annotation. However, existing methods focus solely on clustering unsupervised data while neglecting domain adaptation. Therefore, we propose a consistency-driven prototype-prompting framework for GID from the perspective of integrating old and new knowledge, which includes a prototype-prompting framework for transferring old knowledge from external sources, and a hierarchical consistency constraint for learning new knowledge from target domains. We conducted extensive experiments and the results show that our method significantly outperforms all baseline methods, achieving state-of-the-art results, which strongly demonstrates the effectiveness and generalization of our methods. Our source code is publicly available at https://github.com/smileix/cpp.