[简介] 针对AI大模型训练规模突破万卡、需跨园区长距组网的新需求，提出了一套端到端的联合优化解决方案：分级流控技术将任务完成时间（JCT）降低一个数量级；快速拥塞反馈机制将拥塞反馈延迟从毫秒压缩至数十微秒；逐流负载均衡协议消除端口负载不均问题；采用拓扑亲和的ZeRO 1策略替代ZeRO 3，大幅减少跨园区通信次数。实验网测试表明，通过整合上述关键技术，在千亿参数模型训练场景下可将跨园区训练的算力损失从10%~15%降低至5%以内，验证了长距网络支撑超大规模AI训练的可行性。

[简介]随着人工智能技术的快速发展，传统以比特传输为核心的通信范式正面临根本性变革。本文系统梳理了语义信息处理与通信从理论起源到技术演进的发展脉络，阐明其本质内涵与核心优势，分别从通信与智能两个维度，剖析其商用化面临的关键挑战并研判发展前景。语义通信是6G的重要候选技术，其商用化仍受制于理论体系不完善、全链路技术成熟度不足、标准化进程滞后三重瓶颈；而智能化的持续推进，特别是智能体技术的发展，将有力推动语义通信网的落地部署。前期研究表明，在工业物联网、应急通信、深海水声通信等具有明确语义任务和场景约束的垂直行业中，语义通信技术有望在3～5年内实现局部试商用；伴随6G技术和标准化的演进，语义编解码和语义认知网络架构等有望在5～10年纳入全球6G标准框架；而实现全网络语义内生智能化的大规模部署，则预计需要10～20年。本研究可为学术界与产业界开展语义通信研究及应用部署提供参考。

[Introduction] Fluid antenna systems (FAS) have emerged as a revolutionary technology for next-generation wireless networks, offering continuous spatial degrees of freedom. However, deploying FAS on high-speed mobile embodied intelligent agents induces severe Doppler shifts, leading to an overhead catastrophe when traditional full-port scanning is applied within the sharply reduced channel coherence time. To overcome this bottleneck, this paper proposes a prediction-aided adaptive communication architecture with a local window. Rather than relying on rigid global scanning or highly vulnerable single-point prediction, our mechanism constructs a dynamic local sounding window centered on upper-layer prediction results. By exploiting the inherent spatial channel correlation of FAS, this approach effectively bridges prediction deviations and achieves an optimal balance between pilot overhead and spatial diversity gain. Specifically, we establish a rigorous cross-layer quantitative analysis model for long-term expected throughput. This model integrates a nonlinear bounded hit probability function and a novel gain bridging model based on the zero-order Bessel function of the first kind. Based on this framework, we formulate a discrete optimization problem subject to strict timing constraints to maximize the expected transmission capacity. To solve this non-convex problem efficiently, we design a two-stage low-complexity algorithm that combines offline spatial expectation decoupling with online discrete zeroth-order gradient optimization. This algorithm completely decouples real-time computational complexity from the total number of antenna ports, enabling microsecond-level adaptive tracking of the optimal window. Extensive simulation results demonstrate that the proposed mechanism can adaptively adjust the sounding window scale in response to real-time moving velocity and prediction accuracy. Consequently, it effectively mitigates the pilot overhead catastrophe in highly dynamic scenarios and delivers significantly superior throughput performance compared to traditional fixed-window and global scanning schemes.

[Introduction] This paper investigates joint resource allocation for movable-antenna (MA) -assisted secure integrated data and energy transfer (IDET) systems with passive eavesdroppers. A field-response-based MA channel model is adopted to capture the impact of antenna positions on legitimate, energy-transfer, and eavesdropping links, while energy signals are reused as artificial noise for simultaneous wireless power transfer and secrecy enhancement. A max-min secrecy-rate problem is formulated by jointly optimizing information beamformers, energy beamformers, and MA positions under energy-harvesting, power, movable-region, and antenna-spacing constraints. To address the resulting nonconvexity, an iterative algorithm based on semidefinite relaxation, successive convex approximation, and alternating optimization is developed. Simulation results demonstrate that the proposed scheme improves the minimum secrecy rate compared with fixed-position antenna benchmarks while satisfying the energy requirements of energy receivers.