With the rapid development of information technology, communication systems are increasingly required to support stable, reliable, and real-time information transmission in complex environments. However, during practical communication transmission, signals are often affected by noise interference, channel fading, network jitter, multi-user interference, and packet loss. These problems may reduce speech intelligibility, increase bit error rates, cause video freezing, and weaken the continuity and reliability of communication services. Traditional enhancement techniques rely on stationary assumptions or fixed redundancy, and thus perform poorly under non-stationary noise or dynamically fluctuating network conditions. In recent years, deep learning-based methods have shown strong potential in learning nonlinear mappings from degraded to clean signals, offering better adaptability to diverse distortions. Nevertheless, the analysis shows that AI technology still faces practical deployment challenges, including high computational complexity, heavy dependence on large-scale labeled training data, and limited cross-scenario generalization capability, which require further optimization before such approaches can be reliably integrated into real-time audio/video communication systems.

The rise of generative artificial intelligence has triggered a paradigm shift in the field of video production. This survey systematically examines the evolutionary progress of generative AI video production technologies, with a focus on analyzing the transition from CNN architectures to Transformer-dominated frameworks. Adopting a systematic literature review methodology, this study analyzes relevant academic papers and technical reports. Centered on three core research questions, this survey explores: (1) What architectural innovations have enabled the shift from CNN-based to Transformer-based video generation? (2) What are the current capabilities and limitations of cutting-edge models such as Sora? (3) What challenges and future directions lie ahead for this rapidly evolving field? Key findings demonstrate that the self-attention mechanism of Transformers fundamentally addresses the inherent long-range temporal dependency problem of CNNs, reducing temporal consistency error from approximately 42% for CNNs to roughly 15% for Transformers. Nevertheless, substantial challenges persist in multi-modal consistency, computational resource requirements (approximately 10¹⁸ FLOPs for generating a single 10-second video), and ethical concerns. This survey identifies world models and interactive generation as promising research frontiers for the domain.

Stable cross-domain feature alignment is indispensable for earth observation classification, which is fundamentally hampered by radiometric gaps between generic pre-training images and aerial remote sensing data. Vision-language pre-trained models exhibit strong zero-shot capability on ordinary photos yet suffer severe accuracy loss on satellite and aerial imagery. While LoRA tuning cuts partial training costs, backbone parameter fine-tuning still brings considerable GPU memory overhead in training. Relying on frozen DeepSeek V4 MoE text LLM and static SigLIP vision encoder, this study designs a slim cross-modal projection subnet to eliminate feature distribution gaps between modalities. Stacked residual MLPs constitute the sole learnable part, containing roughly 20M parameters for visual-text latent space matching. The model is evaluated collectively on EuroSAT, PatternNet and RSSCN7, covering nearly 60,000 aerial images with 55 separate scene classes. Recorded aggregate classification precision reached 99.80% across the unified multi-source testing pool. Compared with LoRA-dependent VL-ZSDA-RS benchmark schemes, the adjustable parameter scale shrinks by over half, alongside a 46% cut in peak GPU memory usage. Layer-wise ablation trials reflect unstable matching performance under shallow projection layouts; five stacked transformation layers deliver the most balanced tradeoff between computation overhead and inter-modal alignment quality. External trainable mapping subnetworks, as the test records suggest, unlock visual discrimination capacity for unmodified text-only large language models, supplying a low-hardware threshold tuning route for earth observation research groups constrained by computing resources.

Large language models (LLMs) suffer inherent factual hallucination defects, which block their deployment in high-risk fields such as medicine and finance. Retrieval-Augmented Generation (RAG) serves a mainstream hallucination mitigation solution by introducing traceable external knowledge evidence. Nevertheless, existing RAG variants are plagued by retrieval noise, poor domain generalization, lack of reasoning verification and inconsistent evaluation standards. This paper adopts classification and comparative analysis as core research methods, and systematically sorts out all hallucination suppression technical routes centered on mitigating LLM hallucinations. Four major categories of anti-hallucination RAG technologies are summarized and their applicable boundaries are compared; two mainstream evaluation benchmarks, CRAG and RAGEval, are thoroughly analyzed. Aggregated experimental results demonstrate that layered stacking of multiple technologies achieves optimal hallucination reduction performance. Finally, this paper summarizes existing research gaps, including lightweight deployment and multimodal expansion, and proposes future research directions for trustworthy RAG systems. This review provides systematic theoretical support for industrial RAG model selection and optimization.

The space-air-ground integrated network offers a forward-looking direction for alleviating ground computing power bottlenecks. Efficiently deploying large, distributed models, a key technology for 6G ubiquitous intelligence, in highly dynamic, resource-heterogeneous environments holds significant strategic and application value. This paper systematically reviews the current state of research, core challenges, and key technologies for constructing distributed large-scale models in space-air-ground integrated networks. It analyzes critical issues,, including network heterogeneity, on-board resource constraints, and the security-efficiency trade-off, and evaluates the strengths, limitations, and applicability of existing solutions. The study reveals that adaptive model partitioning, lightweight security protocols, and standardized frameworks remain major shortcomings. Future research should focus on lightweight model design and architecture standardization, deep integration of privacy computing with communication, and joint scheduling of communication, sensing, and computing. This paper clarifies current research limitations and future directions, offering references for the large-scale application of space-air-ground integrated computing power networks.

Neural scene reconstruction has become an important tool for building digital environments used in robotics, autonomous systems, and physical AI training. However, NeRF-based reconstruction often requires high computational cost and does not directly produce simulation-ready mesh assets. 3D Gaussian Splatting offers a faster alternative by representing a captured scene as an explicit cloud of Gaussian primitives that can be rendered in real time. This paper proposes a practical pipeline that converts 3D Gaussian Splatting outputs into mesh-based assets and exports them into NVIDIA Omniverse and Isaac Sim workflows via OpenUSD. The proposed method extracts geometry from the Gaussian primitive cloud, reconstructs a watertight or simulation-usable mesh, bakes appearance information into textures and materials, and exports the result as an OpenUSD-compatible asset with physics and semantic metadata. The goal is to lower the cost of building realistic training environments while preserving visual and geometric fidelity sufficient for simulation, synthetic data generation, and robot learning. The paper also outlines a future extension toward live conversion, where streamed captures are incrementally transformed into simulation-ready scene updates.

As 6G networks advance toward higher levels of autonomy and intelligence, the demand for sophisticated multimodal data processing in communication systems is growing exponentially. Conventional localized AI models encounter significant generalization bottlenecks when handling cross-layer network operations and dynamic resource allocation. To overcome these limitations, this paper systematically investigates the application frameworks of large language models (LLMs) in wireless communication systems—spanning from physical-layer protocols to high-layer network management—while critically evaluating the associated deployment challenges. Drawing on a comprehensive review of prominent literature published over the past three years, this study empirically assesses the performance of diverse LLM architectures across three key domains: physical-layer protocol parsing, network-layer resource allocation, and service orchestration. Results demonstrate that LLMs yield substantial improvements in end-to-end semantic communication, standardized protocol interpretation, and intelligent network resource scheduling. Nevertheless, practical deployment remains severely hindered by the computational constraints of edge devices and prohibitively high inference latency. We conclude that the co-design of lightweight, telecom-specific large language models (Telecom-LLMs) and distributed inference mechanisms constitutes a pivotal evolutionary pathway toward realizing endogenous intelligence in future wireless communication systems.

To address background noise that reduces speech intelligibility in voice communications, this paper compares the noise-reduction performance of FIR (window function), Butterworth IIR, and adaptive LMS filters. Experiments using speech signals with white noise, pink noise, and real-world environmental noise under input SNRs of 0–10 dB show that at 0 dB input SNR, the LMS filter achieves an output SNR of 8 dB and MSE of 0.01, significantly outperforming FIR (6 dB, 0.07) and IIR (5 dB, 0.08). At an input SNR of 10 dB, the performance gap narrows to 1 dB. The IIR filter has the lowest order (5th) and minimal computational cost; the FIR filter (65th) offers the best linear phase; the LMS filter (32nd) exhibits the strongest adaptability at low SNRs. Overall, LMS prioritizes adaptability, IIR prioritizes efficiency, and FIR suits phase-sensitive applications. This study provides quantitative references for noise suppression in voice communication systems.

Event-based visual-inertial odometry (VIO) is often considered a promising sensing solution for low-power micro-robots because event cameras naturally produce sparse, low-latency measurements. However, an efficient sensor does not automatically lead to an energy-efficient system. This paper adopts a system-level view, treating the energy consumption of event-based VIO as an outcome of the interaction between event representation and state-estimation strategy. it examines how information density and computational demand jointly shape system power. From this analysis, it introduces the principle of representation–estimation coupling consistency. Evidence from a broad range of systems supports this principle. The framework also suggests where future low-power VIO systems are likely to converge: event-driven, semi-continuous, and jointly adaptive architectures under strict power budgets. Finally, it provides design guidance for three power ranges and discuss open problems. This study aims to offer a unified perspective for building energy-efficient perception systems for resource-constrained micro-robots.