Multimodal integrated sensing and communication (MMISAC) leverages heterogeneous information sources, such as radio-frequency (RF) signals (e.g., channel state information (CSI) and echoes), vision, radar/light detection and ranging (LiDAR), inertial measurement unit (IMU) measurements, maps, and semantic priors, to enhance environmental awareness and communication efficiency. However, multimodality is not merely about aggregating additional sensors: heterogeneous modalities are inherently misaligned across temporal, spatial, and semantic dimensions, and their outputs must ultimately support predictive and decision-making processes under dynamic network conditions. These challenges also indicate that the future evolution of MMISAC will be increasingly driven by deep learning and artificial intelligence (AI)-native optimization, rather than by isolated improvements in individual sensing or communication modules. In such settings, learning must continuously transform observations into usable representations, anticipate future states, support task-oriented actions, and interpret feedback under resource constraints. Therefore, this survey reviews MMISAC research from a closed-loop perspective, using the information flow of observation, representation, prediction, decision, and feedback as the organizing backbone. Specifically, we systematically examine: (i) multimodal representation learning and fusion strategies that map heterogeneous observations into task-relevant latent states; (ii) predictive modeling paradigms for channel and environmental dynamics, including uncertainty-aware inference; (iii) decision-learning mechanisms that translate predictive knowledge into communication and cooperative actions; and (iv) feedback-driven cross-stage adaptation mechanisms that correct decision policies, prediction models, and latent representations. We discuss evaluation and benchmarking that align communication, perception, and task-level objectives. Finally, we discuss representative applications and future directions for robust and practical MMISAC systems.
Cloud-edge-terminal networks (CETNs) are emerging as a novel paradigm for providing low-latency and energy-efficient computation-intensive services, playing a critical role in enabling latency-sensitive and computation-intensive applications across diverse scenarios. This paper presents a contemporary survey on task offloading in CETNs from the perspectives of latency and energy efficiency. Specifically, a comprehensive overview of task offloading strategies in CETNs is described, tracing the evolution from cloud computing to edge computing and ultimately to collaborative CETN architectures. Task types, offloading modes, and cooperation models (i.e., vertical, horizontal and global offloading) are classified, and their impacts on system performance are analyzed. Key performance indicators such as delay and energy consumption are highlighted, and optimization techniques for enhancing resource management are reviewed. A key finding is that collaborative offloading across cloud, edge and terminal layers significantly outperforms traditional approaches in balancing latency and resource utilization. Yet, achieving optimal coordination remains challenging. We discuss several open research challenges, particularly mobility-induced non-stationarity and proactive adaptability, as well as cross-layer coupling and multi-timescale issues in the cross-layer optimization of communication, computing, and caching (3C) resources, and propose potential future research directions to address these challenges.
Network performance evaluation has long relied on numerical Quality of Service (QoS) and Quality of Experience (QoE) metrics. These quantify network state but cannot explain why a degradation matters or what action should follow. This semantic gap has become a practical bottleneck as networks evolve toward intent-driven and task-oriented architectures. We define semantic network performance evaluation as an evidence-grounded process that interprets network-state or performance evidence against an explicit operational target and produces a contextualized, verifiable judgment, and we distinguish it from adjacent concepts that share its vocabulary or parts of its function. Across the traditional, machine-learning, and large language model (LLM)-based eras, we examine how numerical estimation, explanation, and language-mediated interpretation address different parts of this gap. We classify LLM-enabled methods by three composable descriptors: knowledge adaptation and grounding mechanisms, system-integration patterns, and the LLM’s functional role as a wrapper, reasoner, or surrogate. The LLM-based systems that meet our definition address diagnostic targets. General task-conditioned performance evaluation remains open, and we therefore examine network digital twins and task-oriented evaluation as sources of current evidence and explicit application targets. We synthesize their bounded capabilities in a closed-loop reference architecture that separates semantic inference from numerical calculation and verification. The architecture is a forward-looking synthesis rather than a deployed system. Finally, we analyze six open challenge areas, including benchmarking, reliability, and real-time operation, and propose research directions for each.
Terahertz (THz) communication is increasingly viewed as a promising technology for sixth-generation (6G) wireless networks due to the availability of extremely wide bandwidths supporting ultra-high data rates. At the same time, moving THz communication beyond laboratory prototypes toward practical 6G deployment raises a number of open challenges related to device and circuit technologies, propagation behavior, network design, and spectrum regulation. In this survey, we review the evolution of THz communication with a focus on its role in future 6G systems. We examine various key enabling technologies for THz wireless communications to compare their complementary roles, trade-offs, and technology maturity. We further discuss emerging use cases, mapping each to its dominant requirements in coverage, latency, reliability, hardware complexity, and adaptability. To support deployment-oriented evaluation, the survey also considers key performance indicators (KPIs) and key value indicators (KVIs), linking technical and system-level capabilities to broader prospective value dimensions such as sustainability and societal impact. In addition to physical-layer considerations, we discuss higher-layer implications, highlighting how THz communication links can affect medium-access control, network, and transport-layer design. Furthermore, the paper also explores ongoing global research initiatives, industrial activities, standardization efforts, regulatory developments, and human exposure considerations that shape the emerging THz ecosystem. Finally, we identify open research challenges and outline directions that are critical for the practical realization of THz-enabled 6G networks.
The convergence of integrated sensing and communications (ISAC) with reconfigurable intelligent surfaces (RIS) and metasurface-based architectures is emerging as a key direction for sixth-generation (6G) wireless networks, supporting joint communication and sensing through programmable electromagnetic environments. This survey provides an integrated treatment of metasurface-empowered ISAC across terrestrial, airborne, and spaceborne tiers. Six metasurface architectures are covered: passive and active RIS, simultaneously transmitting and reflecting RIS (STAR-RIS), space-time-coding metasurfaces (STCM), holographic multiple-input multiple-output (MIMO) surfaces (HMIMOS) and the closely related stacked intelligent metasurface (SIM), and metasurface reflectarrays that generate orbital angular momentum (OAM) beams. The survey reviews the electromagnetic foundations, channel models, ISAC waveform design, and performance bounds, then synthesises the terrestrial pillar across beamforming, resource allocation, hardware impairments, and energy efficiency. The OAM-ISAC pillar examines how vortex beams provide additional degrees of freedom for rotational sensing, while critically assessing the practical mode limits and maturity gaps. The non-terrestrial pillar addresses low Earth orbit (LEO) satellites, high-altitude platform stations (HAPS), and unmanned aerial vehicles (UAVs), covering link-budget, Doppler, and handover challenges. Cross-cutting enablers, namely artificial intelligence (AI) driven optimisation and physical-layer security, are surveyed with emphasis on their inter-play with metasurface-ISAC. Two assumptions that pervade the surveyed optimisation literature, perfect instantaneous channel state information and cost-free channel acquisition, are examined critically: a coherence-time-aware accounting of pilot overhead, illustrated by an anchored numerical sensitivity analysis, shows that the reported multiplexing and sensing gains of large surfaces are bounded by the associated acquisition costs. The 3GPP, ETSI, ITU-R, and O-RAN standardisation status frames the route to deployment. Fifteen prioritised open challenges and a technology roadmap to 2035 complete the treatment. By bridging the electromagnetic, signal-processing, and networking communities, this work aims to serve as a reference for researchers and engineers shaping 6G ISAC.
