China Telecom (China)

Visual retrievers such as visual retrieval-augmented generation (RAG) have recently emerged as a powerful model for retrieving multimodal documents without the need to convert page images into text. Existing visual retrievers typically encode every document page separately, ignoring the inherent rich context information across pages within multi-page documents. However, some crucial semantic information often spans multiple pages in a document, and should be effectively encoded for better retrieval. To address this problem, this paper proposes a novel approach utilizing dynamically fusing visual context (DFVC), which adaptively encodes the semantic information across pages. In the proposed DFVC approach, a lightweight plug-and-play adapter is designed; in addition, a contrastive loss function incorporating the positive fused embedding vectors and negative embedding vectors is designed to constrain the adapter, allowing it to learn the weights for the context pages. Together, the designed adapter and loss function allow the retriever to effectively encode useful semantic information across pages while excluding distracting noise. The proposed DFVC is validated on commonly used challenging multi-page document benchmarks. Extensive experimental results demonstrate that it significantly boosts retrieval performance. In addition, the proposed DFVC is highly parameter-efficient since it employs frozen vision-language backbones, allowing it to be easily integrated into existing visual RAG pipelines for finer document retrieval.

Facing the high power and heat density of the AI requirement, this paper reviews recent advances in multi-level thermal performance assessment for data centers. This multi-level framework covering chip, device, rack/row, and room levels. Corresponding to the multi-level framework, a metric system micro-scale and macro-scale indicators is developed to improve system resilience, facility safety and overall energy efficiency. Further, this review summarize progress in data collection, including advanced sensor technologies, computational fluid dynamics (CFD) and reduced-order models, such as AI/ML-based digital twins. Based on recent advances, future research directions are discussed in this paper, including standardization of multi-level metrics, high-accuracy data collection technologies, and utilization of AI/ML based technologies.

Hash center-based deep hashing methods improve upon pairwise or triplet-based approaches by assigning fixed hash centers to each class as learning targets, thereby avoiding the inefficiency of local similarity optimization. However, random center initialization often disregards inter-class semantic relationships. While existing two-stage methods mitigate this by first refining hash centers with semantics and then training the hash function, they introduce additional complexity, computational overhead, and suboptimal performance due to stage-wise discrepancies. To address these limitations, we propose Center-Reassigned Hashing (CRH), an end-to-end framework that dynamically reassigns hash centers from a preset codebook while jointly optimizing the hash function. Unlike previous methods, CRH adapts hash centers to the data distribution without explicit center optimization phases, enabling seamless integration of semantic relationships into the learning process. Furthermore, a multi-head mechanism enhances the representational capacity of hash centers, capturing richer semantic structures. Extensive experiments on three benchmarks demonstrate that CRH learns semantically meaningful hash centers and outperforms state-of-the-art deep hashing methods in retrieval tasks.

We present PHOTONS (Pose-Free Human-Centric Photo-Realistic Real-Time Novel View Synthesis from Sparse Views), a real-time framework for novel view synthesis without requiring camera calibration. Our method reconstructs consistent 3D Gaussian point clouds and synthesizes 2K photo-realistic novel views from arbitrary numbers (>=2) of freely placed cameras. PHOTONS faithfully renders dynamic human bodies amid complex backgrounds, including interactive object manipulation and fine-grained details (e.g., hair strands), while maintaining 25 FPS throughput on commodity GPU like NVIDIA RTX 4090. By combining pose-free spatial point cloud reconstruction with Gaussian parameter estimation, our method demonstrates strong resilience to occlusions and camera perturbations. Additionally, we develop a 3D stereo system that drastically reduces setup complexity compared to existing solutions. Experiments on public and custom datasets show that PHOTONS outperforms state-of-the-art methods in both efficiency and visual quality.

Waterway transport is a cornerstone of green, low-carbon logistics, yet it is increasingly vulnerable to multi-hazards — ranging from extreme weather and flooding to infrastructure failures and traffic congestion. To address these threats, the project "Internet of Ships" (船联网) was initiated 10 years ago across the Yangtze River Delta and the Grand Canal. This project aims to transform traditional shipping into a highly resilient, intelligent system that connects ships, the shore, and land-based management centers to one another. To ensure this system could withstand real-world systemic shocks, the project team analyzed over 5,000 surveys and conducted 120 field studies. This research led to a systematic strategy that directly enhances the waterway resilience: 1. Multi-Hazard Sensing on infrastructure and vessels for early warnings before a minor issue turns into a systemic failure. 2. Data Fusion for Rapid Response in case of emergency to avoid cascading effects through the network. 3. Smart Emergency Coordination to rescue and re-routes traffic and to retain the network functions even under stress. 4. Operational Efficiency is improved by establishing the Waterway ETC (non-stop lock passage) that alone saves 590 million RMB annually by reducing idling time — making the system more efficient and less prone to the cascading delays often seen in multi-hazard events. Ultimately, this research demonstrates that digital soft infrastructure can reinforce physical hard infrastructure, creating a shipping network that is not only green but robust enough to recover quickly from diverse environmental and operational hazards.

Anonymous communication networks such as the Invisible Internet Project (I2P) are essential for safeguarding privacy and ensuring freedom of expression, necessitating robust performance and security evaluation in controlled environments. Network testbeds offer a reliable alternative to real-world testing. This paper proposes a dynamic modeling framework based on Graph Attention Network (GAT). We introduce a Region-Centric Initialization (RCI) strategy to establish an initial observation anchor, followed by a GAT-based Locality-Aware (GAT-LA) sampling mechanism that treats representative node selection as a dynamic learning task. Experimental results demonstrate that the GAT-LA mechanism significantly outperforms static methods in maintaining long-term similarity to real-world I2P performance metrics. The integrated stability penalty mechanism effectively suppresses excessive topological fluctuations, ensuring temporal smoothness across evolutionary cycles. Furthermore, the RCI strategy provides high engineering flexibility by supporting both automated scoring and target-oriented manual configuration. This paper presents a scalable methodology for dynamic network simulation with enhanced statistical alignment, providing a practical reference for security research within resource-constrained anonymous network ranges or testbeds.

Asynchronous Federated Learning (AFL) has emerged as a promising paradigm to address the challenges posed by heterogeneous device environments in federated learning systems. However, the problem of low accuracy and slow convergence due to inconsistent updates from stragglers and normal clients remains severe in AFL. Previous works either discard or penalize the updates from stragglers, which can lead to the loss of valuable data or introduce bias into the model. Furthermore, existing AFL frameworks integrating synchronous optimization algorithms face the challenges of weak compatibility and scalability, limiting large-scale training. In this paper, we propose <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DVAFL</i>1 an efficient and scalable AFL framework that significantly improves the performance of AFL in terms of model accuracy and convergence speed by effectively utilizing and compensating for the updates from stragglers, while naturally integrating synchronous optimization algorithms. Specifically, DVAFL introduces a dynamic window protocol for adaptive aggregation to balance the contribution of stragglers and ensure faster and more stable convergence. Further, the version control mechanism corrects stale gradients by compensating for the missed model updates of stragglers, thereby improving model performance. Extensive experiments on three public datasets demonstrate that DVAFL achieves an average convergence speed 2.3× faster and an accuracy improvement of 5.5% compared to state-of-the-art AFL methods.

Recent progress in image super-resolution (SR) has achieved impressive reconstruction quality, but most existing approaches rely on complex architectures with large parameter counts and heavy computational costs. In this work, we propose a lightweight yet powerful framework termed Dual Adaptive Feature Modulation (DAFM) for image super-resolution. The key idea is to integrate spatial and channel adaptive modulation mechanisms to jointly enhance feature representation capability. Specifically, the Spatially-Adaptive Feature Modulation focuses on capturing and integrating local structural details by adaptively emphasizing informative spatial regions. In contrast, the Channel-Adaptive Feature Modulation aggregates both intra-channel and global contextual information to enhance feature interaction and representation capability. By effectively combining dual adaptivity within a concise design, DAFM achieves strong reconstruction performance with remarkably fewer parameters and lower FLOPs compared to recent lightweight SR models. Extensive experiments demonstrate that our method yields high image quality without compromising model compactness.