On-Policy Self-Distillation (OPSD) uses privileged information available only to the teacher to provide dense token-level supervision on trajectories generated by the student. However, existing methods often rely on verified solution traces, explanations generated by external models, or manually localized visual evidence, which limits their scalable application to multimodal large language models. To address this issue, we exploit the information gap between high- and low-resolution views of the same image and propose RP-OPSD (Resolution-Privileged On-Policy Self-Distillation for Multimodal Large Language Models). During training, the student policy generates on-policy trajectories from images at one-quarter of the original resolution, while the teacher policy provides supervision using the original-resolution images. By minimizing the divergence between their output distributions along the student trajectories, the student learns the predictive behavior of the teacher under high-resolution inputs, thereby strengthening its low-resolution capability and transferring the learned improvement to original-resolution inference. RP-OPSD requires neither additional human annotations nor external models to generate solution traces but only image--question pairs. Experiments on Qwen3.5-9B show that RP-OPSD achieves a 5.45\% relative improvement in average performance at the original resolution and a $1.78\times$ training speedup over OPSD. These results demonstrate that resolution differences can serve as a simple and scalable source of privileged information, providing an effective and efficient approach to on-policy self-distillation for multimodal large language models.

Vehicular Ad-hoc Networks (VANETs) are the digital cornerstone of autonomous driving, yet they suffer from severe network fragmentation in urban environments due to physical obstructions. Unmanned Aerial Vehicles (UAVs), with their high mobility, have emerged as a vital solution to bridge these connectivity gaps. However, traditional Deep Reinforcement Learning (DRL)-based UAV deployment strategies lack semantic understanding of road topology, often resulting in blind exploration and sample inefficiency. By contrast, Large Language Models (LLMs) possess powerful reasoning capabilities capable of identifying topological importance, though applying them to control tasks remains challenging. To address this, we propose the Semantic-Augmented DRL (SA-DRL) framework. Firstly, we propose a fragmentation quantification method based on Road Topology Graphs (RTG) and Dual Connected Graphs (DCG). Subsequently, we design a four-stage pipeline to transform a general-purpose LLM into a domain-specific topology expert. Finally, we propose the Semantic-Augmented PPO (SA-PPO) algorithm, which employs a Logit Fusion mechanism to inject the LLM's semantic reasoning directly into the policy as a prior, effectively guiding the agent toward critical intersections. Extensive high-fidelity simulations demonstrate that SA-PPO achieves state-of-the-art performance with remarkable efficiency, reaching baseline performance levels using only 26.6

The adaptability and lightweight decision-making capabilities of Multi-Agent Reinforcement Learning (MARL) has been widely applied in multiple Unmanned Aerial Vehicles (UAVs) assisted communication for trajectory planning. However, in multi-hotspot scenarios, existing MARL based algorithms suffer from sparse rewards and inefficient credit assignment, leading to poor learning efficiency in UAV agents and constraining the energy efficiency. Notably, Large Language Models (LLMs), leveraging their vast prior knowledge and powerful reasoning capabilities, have demonstrated great potential for optimizing complex systems. However, it is not trivial to deploy LLMs on UAVs with low computational resources. Therefore, we propose a novel LLM-enhanced Multi-Agent Proximal Policy Optimization algorithm (LLM-MAPPO) for multi-UAV trajectory planning. LLM-MAPPO integrates the reasoning advantages of LLMs with the lightweight decision-making capabilities of MAPPO agents, significantly enhancing the global energy efficiency of multi-UAV trajectories while ensuring deployability on UAVs. In LLM-MAPPO, we first design a Chain-of-Thought (CoT) prompt to guide the LLM in reasoning and generate high-quality expert trajectory policies with superior credit assignment capabilities. Then, we develop a dual-phase update algorithm that enables MAPPO agents to inherit knowledge from LLM expert policies and achieve further surpassing. This algorithm enhances credit assignment effectiveness and addresses the sparse reward problem to improve learning efficiency. Experiments show that LLM-MAPPO accelerates agent training by 67% and outperforms compared algorithms among various metrics, notably improving energy efficiency by 194%.

Practical NV-center magnetometry requires both accurate modeling of open quantum dynamics and robust control under environmental drift. However, commonly used perturbative descriptions inadequately capture dissipative processes, while existing control strategies remain individually limited: open-loop protocols cannot adapt to unknown frequency variations, Lyapunov feedback typically converges only to a neighborhood of the optimal sensing state, and reinforcement learning alone lacks stability guarantees and requires extensive training. Here we develop a Lindblad-based framework for one effective NV center that explicitly incorporates relaxation and dephasing; the analytical model does not describe a collective many-NV ensemble. We introduce a dissipation-triggered feedback mechanism that restores high-sensitivity states following spontaneous emission events. Building on this model, we propose a hybrid control architecture that combines Lyapunov stabilization with reinforcement-learning-based adaptive correction to compensate unknown Hamiltonian drift. Robustness tests over detuning and microwave-amplitude variations show that Hybrid attains the highest QFI, substantially lower cross-realization dispersion than Pure RL and Robust GRAPE, faster fixed-budget learning than Pure RL, and substantially lower offline dynamical-propagation cost with smoother control modulation than Robust GRAPE. This physics-informed and data-driven strategy provides a robust and resource-efficient approach for maintaining high quantum Fisher information in realistic NV sensing environments.

With the increasing application of Unmanned Aerial Vehicles (UAVs) in urban areas, employing ground Base Stations (BSs) to serve UAVs has been proposed as a low-cost and promising solution. However, the Quality of Service (QoS) of UAVs served by ground BSs is impacted by multiple factors, including blockage, UAV height, transmit power, BS density, BS selection strategy, and so on. Efficient and accurate evaluation of the communication performance between UAVs and BSs is critical for network planning and optimization. In this paper, we analyze the communication performance of UAVs served by ground BSs in urban environments, where millimeter wave communication is employed. In contrast with existing works adopting a simplified Line-of-Sight (LOS) probability model, we adopt a more practical yet complex LOS model proposed by 3GPP for characterizing the densely distributed buildings in urban environments. Moreover, the distribution of BSs is modeled as a Mat & eacute;rn Hard-Core Point Process (MHCPP) with a minimum distance constraint to reflect real scenarios. The analytical expressions of outage probability and ergodic capacity under different BS selection strategies are derived for enabling efficient performance evaluation. We verify the accuracy of the analytical results through simulation experiments under both the Urban Macro (UMa) and Urban Micro (UMi) scenarios, while the impacts of multiple parameters are analyzed. The results indicate that the outage probability decreases as the UAV height increases, but the ergodic capacity shows an opposite trend. Moreover, we show that our analytical results can be used to select the optimal flight height for UAVs with a given required outage probability.

Table question answering (TableQA) is a fundamental task in natural language processing (NLP). The strong reasoning capabilities of large language models (LLMs) have brought significant advances in this field. However, as real-world applications involve increasingly complex questions and larger tables, substantial noisy data is introduced, which severely degrades reasoning performance. To address this challenge, we focus on improving two core capabilities: Relevance Filtering, which identifies and retains information truly relevant to reasoning, and Table Pruning, which reduces table size while preserving essential content. Based on these principles, we propose EnoTab, a dual denoising framework for complex questions and large-scale tables. Specifically, we first perform Evidence-based Question Denoising by decomposing the question into minimal semantic units and filtering out those irrelevant to answer reasoning based on consistency and usability criteria. Then, we propose Evidence Tree-guided Table Denoising, which constructs an explicit and transparent table pruning path to remove irrelevant data step by step. At each pruning step, we observe the intermediate state of the table and apply a post-order node rollback mechanism to handle abnormal table states, ultimately producing a highly reliable sub-table for final answer reasoning. Finally, extensive experiments show that EnoTab achieves outstanding performance on TableQA tasks with complex questions and large-scale tables, confirming its effectiveness.

While model predictive control (MPC) effectively handles constrained quantum dynamics, existing methods often lack rigorous guarantees for recursive feasibility and stability. This paper proposes a Lyapunov terminal cost-assisted quantum model predictive control (LTC-QMPC) strategy for two-level quantum systems, where we use a Lyapunov function as the terminal cost within the MPC framework and construct control sequences subject to optimization constraints via the terminal Lyapunov control law. This approach establishes the recursive feasibility of LTC-QMPC and achieves the asymptotic stability of the controlled quantum system. Numerical results show that LTC-QMPC achieves faster convergence to a target eigenstate of the internal Hamiltonian compared with standard and finitetime Lyapunov controls. Furthermore, the LTC-QMPC control law exhibits strong robustness against uncertainties in the Hamiltonian.

This work presents a comprehensive framework for reliable multi-qubit teleportation in a thousand-node quantum network, addressing key challenges such as resource contention, fidelity degradation, and stochastic entanglement distribution. We introduce a multi-server spatio-temporal queue model to dynamically manage repeater node states, enhancing path discovery efficiency by accurately estimating the waiting time and resource availability. To maintain high fidelity over long distances, we propose a hop-by-hop backup mechanism for end-to-end entanglement purification, decomposing fidelity requirements into individual link constraints and employing multi-round purification with backups. Additionally, we develop a deterministic scheduling scheme that decomposes complex multi-qubit requests into simpler one-qubit paths, ensuring reliable transmission with high success probability. Based on NetSquid’s simulation of quantum hardware, our framework is implemented on a real-world quantum communication network protocol stack and tested on a platform with 1024 repeater nodes, demonstrating significant improvements in repeater latency, end-to-end fidelity and teleportation reliability of large-scale quantum networks.

Microservice architecture offers great scalability and flexibility to the development of online services systems. Performance anomalies, which happen frequently due to code bugs or runtime environment misconfiguration, can severely damage the system availability and cause great losses. However, it is challenging to detect performance anomalies and locate their root causes considering the large volume of monitoring data (e.g., metrics and traces) and the complex dynamic interdependence between heterogeneous services. Against these challenges, we propose HG-PAD, an automatic performance anomaly diagnosis (PAD) framework for microservice systems. We build the multi-relation heterogeneous graph to model the intricate dependency between services. We further design a structure learning mechanism combining graph neural network (GNN) and node embedding learning to capture the dynamic and latent dependencies. Based on the optimized dependency graph, we devise a Conditional Variational Auto-Encoder (CVAE) based unsupervised anomaly detection method and a graph attention network (GAT) based root cause localization method for accurate anomaly diagnosis. We use datasets of different scales based on real world applications to verify the effectiveness of HG-PAD, and the experimental results show that HG-PAD achieves better diagnostic performance compared with existing baseline methods.

Space-Ground Integrated Networks (SGINs) hold the potential to enable seamless and high-quality global coverage in an economically viable manner. However, the limited bandwidth and relatively long delay of satellite-ground links pose significant challenges in meeting the increasing demands driven by surging traffic. To address this issue, we propose a three-tier cooperative caching architecture incorporating base stations (BSs), satellites, and a content server, which aims to reduce the average content retrieval delay through cooperative caching between BSs and satellites. The joint optimization problem among them is both challenging to solve and non-scalable due to its exponentially increasing computational complexity. Meanwhile, the geographic characteristics of requests are overlooked in existing work. As a consequence, we propose a novel block partitioning algorithm based on the popularity similarity across areas to facilitate cooperative caching, which reduces the computational complexity, and ensures the scalability and the effectiveness of cooperative caching. Based on the partitioned blocks, the optimization problem is decomposed into two subproblems: intra-block cooperative optimization among BSs and inter-block cooperative optimization among satellites, effectively catering to the characteristics of wide-area coverage in SGINs. For the subproblems with finite dimensions, Semidefinite Relaxation (SDR) based intra-block and inter-block cooperative caching approaches are proposed to obtain the optimal cooperative caching strategies. Extensive simulations demonstrate that, compared to directly solving the original problem, the proposed algorithm reduces the solving time from exponential to linear growth. Moreover, our algorithm outperforms existing schemes by reducing 13% average retrieval delay and improving 12% overall cache hit rate.

Website Fingerprinting (WF) attacks have posed a serious threat to the anonymity of the onion router (Tor) communication system, as attackers can passively pry into the encrypted traffic and infer the website visited by users. To defend against WF, recent studies focus on adversarial perturbations. However, most of them suffer from a high bandwidth overhead and a low defense performance. To address this problem, our basic idea is to generate perturbation only on the sensitive regions, which can effectively mask the website’s fingerprint, thus misleading the WF attack models and reducing the bandwidth overhead. In this paper, we formulate a joint optimization problem of perturbation position and magnitude by confining the perturbations within sensitive regions, which is rarely considered in the literature. We propose a robust low-overhead WF defense method based on reinforcement learning (RL), named RLpatch. RLpatch identifies the common sensitive regions of various surrogate models and adjusts perturbation according to the query result from a query WF model. It further employs the positional frequency of perturbations to generate a common perturbation paradigm for different traces of a same website. Experimental results show that RLpatch achieves higher defense performance, lower bandwidth overhead and better robustness against adversarial training compared to the state-of-the-art methods.

Pipeline parallelism is an efficient approach for distributed deep neural network (DNN) model training. However, existing pipeline parallel training systems often employ fixed model allocation strategies, leading to suboptimal efficiency in dynamic training scenarios. To address this issue, this paper proposes an efficient adaptive pipeline training system, iPipe, which dynamically manages the stages of the DNN model on GPUs based on their real-time states, effectively enhancing the training efficiency. In iPipe, a model allocation algorithm based on dynamic programming method is designed to search for an optimal model allocation plan with low complexity. Additionally, a live migration strategy which adopts an alternating computation method is proposed to achieve dynamic migration of model parameters between workers with minimal data transfer. We evaluate iPipe on a real GPU cluster, and the experimental results demonstrate that iPipe outperforms other pipeline parallel training systems in terms of overall training time for DNN models.

In this paper, an optimal asymmetric controlled teleportation protocol is proposed, where a three-dimensional (3D) GHZ entangled state is utilized to teleport an arbitrary unknown two-dimensional (2D) qubit and the correlation between the two entangled qutrits caused by their continuous transmission through a noisy channel is considered. We design a high-dimensional feed-forward control operator and use weak measurement reversal instead of the unitary operations in standard teleportation to mitigate the reduction in fidelity caused by noise. We further derive the average fidelity and overall success probability of the proposed teleportation protocol, and simulate its performance under correlated amplitude damping (CAD) and amplitude damping (AD) noise channels. The simulation results show that our protocol significantly improves the average fidelity under CAD noise. In particular, under AD noise, the average fidelity remains constant at 1. Moreover, we derive the optimal overall success probability without compromising fidelity. To verify the superiority of the feed-forward control method, we also calculate and analyze the performance of an asymmetric controlled teleportation protocol using only environment-assisted measurement (EAM), and compare it with the performance of our protocol through simulation. The results demonstrate that our protocol outperforms the protocol using only EAM methods in both CAD and AD noise.

In recent years, quantum communication and quantum computing have become strategic technologies of great importance to countries around the world, with profound implications for individuals, businesses, and even nations. Among these, graph states are a very important class of multi-particle entangled states, playing a crucial role in fields such as quantum secret sharing, measurement-based quantum computing, and quantum metrology. However, the distribution of graph states faces the following challenges: existing quantum communication networks designed for single-qubit transmission cannot effectively accomplish graph state distribution tasks, and current methods for graph state distribution lack consideration of node resources, making it difficult to complete graph state distribution in some cases.We conduct research on cross-node graph states distribution methods in quantum communication networks under limited resource constraints. To address the issue that current quantum communication networks cannot handle graph states and other multi-particle entangled states, a graph state distribution method is proposed. This method combines local operations and classical communication to meet the distribution requirements for graph states, which are multi-particle entangled states. Additionally, to solve the problem of limited quantum resources at individual nodes, we proposed a graph states segmentation method. This method divides a complete graph state into subgraphs for distribution and subsequently integrates them, enabling graph state distribution services in a broader range of resource scenarios.

With the rapid development of Unmanned Aerial Vehicle (UAV) technology, homogeneous UAV swarms have become a core force in post-disaster reconnaissance tasks. Their key requirement is to achieve efficient coverage of scattered targets while satisfying operational constraints. However, traditional path planning methods exhibit significant limitations in simultaneously addressing target coverage, flight endurance constraints, and swarm load balancing. To address these issues, we propose a two-stage cooperative task planning method for multi-UAV reconnaissance. The first stage employs a coverageaware clustering algorithm, which groups distributed task points according to the UAVs' sensing range to ensure reasonable task assignment and reduce computational complexity. The second stage utilizes a path planning algorithm founded on the Simulated Annealing-enhanced Genetic Algorithm (SAGA), which optimizes task assignment under flight distance constraints by minimizing the total flight distance and balancing the workload in the swarm. Extensive evaluations demonstrate that our proposed method achieves approximately 40% faster convergence speed, along with nearly 30% reduction in both total path length and task completion time.

Robust Privacy-Preserving Federated Learning (PPFL) is a secure distributed machine learning paradigm designed for untrusted environments, aiming to aggregate gradients while ensuring the reliability of the results without disclosing user gradients. However, existing single-server robust PPFL schemes require users to generate commitments in each aggregation round to ensure the correctness of the aggregation results, which leads to high computational overhead. We propose an efficient single-server robust PPFL scheme named Privacy-Preserving Robust Federated Learning (PRFL). PRFL achieves efficient gradient aggregation through a “detection-identification-exclusion" strategy. PRFL only performs quick detection without requiring high-complexity commitments in most of aggregation round, thereby ensuring excellent efficiency. PRFL comprises three pivotal components: Privacy-Preserving Gradient Aggregation Based on Packed Secret Sharing (PGAPS), Swift Share Verification based on Dual Codes (SSVDC), and Commitment-based Malicious User Identification (CMUI). PGAPS is utilized to implement the FLTrust rule without disclosing gradients. SSVDC swiftly detects incorrect shares without using commitments. CMUI identifies malicious users when SSVDC detects incorrect shares. Experimental results demonstrate the robustness and efficiency of PRFL. In a PPFL system with 100 users, PRFL can robustly aggregate gradients of a million dimensions within 37 seconds of average computational time.

In this paper, we study the cooperative autonomous inspection by Unmanned Aerial Vehicles (UAV) swarms in dynamic and uncertain environments. Facing the challenges of spatio-temporally random target events and constrained onboard resources, we propose a distributed online path planning algorithm based on a Hybrid Artificial Potential Field (HAPF). In contrast to existing works relying on a single probability map, we construct a hybrid artificial potential field to integrate a probability map and an exploration map. The probability map, generated via Kernel Density Estimation from historical and real-time data, learns the distribution pattern of events, while the exploration map encourages coverage of unvisited areas. Virtual attractive forces guide UAVs to prioritize high-probability events and explore unknown regions, whereas virtual repulsive forces ensure collision-free cooperation. Moreover, a Hungarian algorithm based dynamic task allocation method is proposed to prevent resource redundancy. To tackle the well-known local minimum problem in potential field, a sticking release mechanism is designed to escape local optima. Simulation results demonstrate that, in comparison to existing algorithms, our method improving the task completion rate by 20 % and shortening the average completion time by 13 %, offering a practical solution for resource-constrained UAV swarms to achieve efficient and safe autonomous inspection in complex and dynamic environments.

Correlating individual alerts to reconstruct attack scenarios has become a critical issue in identifying multi-step attack paths. Most of existing reconstruction approaches depend on external expertise, such as attack templates or attack graphs, to identify known attack patterns, which are incapable of uncovering unknown attack patterns that exceed prior knowledge. Recently, several expertise-independent methods utilize alert similarity or statistical correlations to reconstruct multi-step attacks. However, these methods often miss rare but high-risk events. The key to overcoming these drawbacks lies in discovering the potential causalities between security alerts. In this paper, we propose GRAIN, a novel graph neural network and reinforcement learning aided causality discovery approach for multi-step attack scenario reconstruction, which does not rely on any external expertise or prior knowledge. By matching the similarity between alerts’ attack semantics, we first remove redundant alerts to alleviate alert fatigue. Then, we correlate these alerts as alert causal graphs that embody the causalities between attack incidents via causality discovery. Afterwards, we employ a graph neural network to evaluate the causal effect between correlated alerts. In light of the fact that the alerts triggered by multi-step attacks have the maximum causal effect, we utilize reinforcement learning to screen out authentic causal relationships. Extensive evaluations on 4 public multi-step attack datasets demonstrate that GRAIN significantly outperforms existing methods in terms of accuracy and efficiency, providing a robust solution for identifying and analyzing sophisticated multi-step attacks.

Distributed training of deep neural networks (DNNs) suffers from efficiency declines in dynamic heterogeneous environments, due to the resource wastage brought by the straggler problem in data parallelism (DP) and pipeline bubbles in model parallelism (MP). Additionally, the limited resource availability requires a trade-off between training performance and long-term costs, particularly in online settings. To address these challenges, this article presents a novel online approach to maximize long-term training efficiency in heterogeneous environments through uneven data assignment and communication-aware model partitioning. A group-based hierarchical architecture combining DP and MP is developed to balance discrepant computation and communication capabilities, and offer a flexible parallel mechanism. In order to jointly optimize the performance and long-term cost of the online DL training process, we formulate this problem as a stochastic optimization with time-averaged constraints. By utilizing Lyapunov's stochastic network optimization theory, we decompose it into several instantaneous sub-optimizations, and devise an effective online solution to address them based on tentative searching and linear solving. We have implemented a prototype system and evaluated the effectiveness of our solution based on realistic experiments, reducing batch training time by up to 68.59% over state-of-the-art methods.