Large Language Models (LLMs) face significant memory and latency overheads during long-context inference due to the growing KV cache, especially in Knowledge Base Question Answering (KBQA) settings that require support for multiple downstream queries. Query-aware eviction methods do not generalize across queries, while existing query-agnostic approaches rely on a single proxy query, leading to fragile eviction decisions under high eviction ratios. We propose ContrastKV, a robust query-agnostic KV cache eviction algorithm for multi-query generalization. ContrastKV introduces a contrastive signal fusion mechanism that jointly exploits complementary semantic and non-semantic signals. By contrasting semantic consistency with structural robustness, the method constructs a more reliable eviction criterion that alleviates the blind spots of single-query proxies. The framework integrates efficient signal generation, parallel importance scoring, and multi-level fusion across heads and layers. Experiments show that ContrastKV outperforms state-of-the-art methods, retaining up to 92% accuracy with only 20% of the KV cache budget, while reducing decoding latency by approximately 50% and significantly lowering GPU memory usage.

The progress in Natural Language Processing (NLP) using Large Language Models (LLMs) has greatly improved medical sentiment analysis of patient feedback extraction from health-related question and answer. However, using LLMs to analyze such data often requires significant training data and computational resources, resulting in considerable increases in training costs and durations, which is one of the primary issues in applying LLMs to real-world healthcare scenarios. To tackle these challenges, a novel optimization algorithm named KAdam-EnGPT4LLM, based on Kalman filters and Adaptive Moment Estimation, is proposed to enhance training efficiency and reduce training costs of LLMs for analyzing patient feedback sentiment. Furthermore, the optimization algorithm KAdam-EnGPT4LLM is employed in training the LLM model GPT4ALL for medical sentiment analysis, resulting in the development of GPT4ALL-MediSentAly-KAdam, which leds to faster convergence and more stable training specifically for medical questions and answer in the context of healthcare. The results show that our GPT4ALL-MediSentAly-KAdam with the optimization algorithm KAdam-EnGPT4LLM achieved better performance that include the best Accuracy, Recall, F1-score, and Runtime for both datasets, outperforming traditional fine-tuned LLMs such as the classic GPT4ALL, Ada, Babbage, Curie, and Duvinci.

In the 6G-enabled intelligent transportation systems (ITS), each intelligent transportation terminal needs to perform long-distance, low-latency image interaction to ensure real-time information exchange, including real-time vehicular environmental images and various vehicular media images. However, due to high computational cost and large computing resource usage, many learning-driven image compression models are difficult to deploy on intelligent transportation terminals such as edge devices and connected vehicle terminals in the 6G-enabled ITS to reduce the transmission resource consumption of massive image data from ITS. To address the above problems, this paper proposes a high-throughput lightweight Transformer model for image compression tasks on ITS-based intelligent transportation terminals. By constructing a lightweight Transformer combination and reducing the use of Transformer blocks by internally connecting convolutional blocks, the computational cost of the model is reduced. Furthermore, this paper proposes a lightweight Swin Transformer module, which further reduces the resources required for model calculation by directly connecting window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SW-MSA). In addition, this paper designs a simplified entropy model, which speeds up the execution time of the entropy model and improves the throughput of the overall network model by directly fusing the latent representation of the images. Experimental results show that our proposed model has significantly improved throughput and execution time compared with the state-of-the-art models on two ITS public datasets.

With the continuous improvement of device imaging resolution, the popularity of Ultra-High-Definition (UHD) images is increasing. Unfortunately, existing methods for fusing multi-exposure images in dynamic scenes are designed for low-resolution images, which makes them inefficient for generating high-quality UHD images on a resource-constrained device. To alleviate the limitations of extremely long-sequence inputs, inspired by the Large Language Model (LLM) for processing infinitely long texts, we propose a novel learning paradigm to achieve UHD multi-exposure dynamic scene image fusion on a single consumer-grade GPU, named Infinite Pixel Learning (IPL). The design of our approach comes from three key components: The first step is to slice the input sequences to relieve the pressure generated by the model processing the data stream; Second, we develop an attention cache technique, which is similar to the KV cache for infinite data stream processing; Finally, we design a method for attention cache compression to alleviate the storage burden of the cache on the device. In addition, we provide a new UHD benchmark to evaluate the effectiveness of our method. Extensive experimental results show that our method maintains high-quality visual performance while fusing UHD dynamic multi-exposure images in real-time (>40fps) on a single consumer-grade GPU.

Beyond 5G-enabled edge computing networking (ECN) will further deploy computing and communication resources to the edge of the networks. Then, edge service demands for Internet of Things (IoT) applications are becoming more and more diverse, while the corresponding routing service capability is limited and not flexible enough to deal with the demands of ECN, which then leads to reducing the inherent routing capability of ECN. It becomes extremely difficult for ECN to support diversified demands and provide diverse IoT applications quickly and flexibly. In this article, we propose a novel and customized deep routing mechanism for IoT applications in ECN, in which the network slicing and deep learning methods are jointly applied and leveraged. First, we design a new ECN architecture that formulates four kinds of network slices to cope with various IoT scenarios, which are eMBB, uRLLC, mMTTC, and backup slices. Second, using these slices, we can customize the ECN environment flexibly, based on which we propose the corresponding routing method for the purpose of fast and efficient service delivery. In particular, the mapping between network slices and the infrastructure is established with the object of maximizing the resource utilization. Then, the routing is designed and customized by using the deep learning model. Lastly, the experimental results show that the deep customized mechanism designed in this article can reduce the average loss rate of the model, decrease the average delay, as well as improve the average resource utilization compared with the existing studies.

In the context of vehicle-road-cloud integration, multi-vehicle platooning systems have become an important approach for improving road traffic efficiency, driver comfort, driving safety, energy consumption, and mitigating traffic congestion. However, under high-speed mobility, Vehicle-to-Vehicle (V2V) communication within multi-vehicle platoons is susceptible to delays caused by interference and the inherent uncertainties of wireless communication channels. These delays present considerable challenges to achieving effective multi-vehicle cooperative control. To overcome the limitations of existing research, this paper proposes a joint communication and control optimization strategy for multi-vehicle platooning systems. A novel spacing error metric is introduced, which uses the real-time velocity of each vehicle to improve the platooning system responsiveness. Furthermore, we derive the Signal-to-Interference-plus-Noise Ratio (SINR) threshold to ensure the stability and reliability of the platoon. This ensures safe distances and synchronized speeds among all vehicles, even when communication delays occur. Finally, the proposed joint optimization strategy is validated through performance comparisons, demonstrating its effectiveness and superior performance.

In the era of large-scale machine learning, largescale clusters are extensively used for data processing jobs. However, the state-of-the-art heuristic-based and Deep Rein-forcement Learning (DRL) based job scheduling mechanisms are facing challenges such as slow training speed and underexploitation of jobs' complex dependencies. We propose MPDA, a Massively Parallel learning and Dependency-Aware scheduling algorithm, consisting of a fast-training mechanism and a novel dependency-aware policy network, GATNetwork, to address these two challenges respectively. The fast-training mechanism is a two-level massively parallel training method that can significantly accelerate the training process and maximally utilize the resources of the cluster. Additionally, its decoupled learning and interacting design enables hybrid-workload training for MPDA, which guarantees the generalization and robustness of MPDA. The GATNetwork exploits the dependencies among stages/jobs using Graph Attention Network (GAT) and Long Short-Term Memory (LSTM) networks to improve the performance of the scheduling policy. The experiments show that MPDA accelerates the training speed by one to two orders of magnitude and achieves better scheduling performance, i.e., lower average job completion time, compared with existing scheduling algorithms.