Edge-deployed wireless foundation models are accessed as compact discrete token streams, an emerging AI-native 6G workload. Such tokens carry heterogeneous perceptual importance to which the radio access network is blind. The gap is most costly over Non-Terrestrial Networks (NTN). Using Mimi, the streaming neural tokenizer of the Moshi speech foundation model, we first show that adding the first residual VQ layer to the semantic base layer reduces the Word Error Rate (WER) from 94% to 32%, establishing a dual semantic/structural token hierarchy. We exploit this hierarchy in the Semantics-Aware Cross-Layer Framework (SACF): a 16-bit token-importance header reaches the gNB through standards-aligned signaling, an existing GTP-U extension header (3GPP TS 29.281) and a proposed lightweight SDAP extension (TS 37.324). There, a Semantic Reliability Control (SRC) policy module drives semantic link adaptation, HARQ management, and LLR-driven error tolerance at the MAC layer, without retraining the codec or modifying the physical layer. The abstraction is codec-agnostic by design, though validated here on Mimi alone. Because unequal protection matters only when the resource cannot protect everything, the evaluation is rate constrained: over standards-anchored geostationary NTN channels with a 5G NR LDPC physical layer, the scheduler grants a fixed number of coded bits per 80 ms frame, shared by every scheme. At $-1$ dB link SNR and a 320-bit grant, exposing the token layer structure to the scheduler cuts WER from 0.70 to 0.40 at equal consumed resource, a saving equivalent to 35.9% of the bits the semantics-blind baseline must spend to match it. Ablation attributes the effect to that exposure rather than to the sophistication of the policy built on it: grading protection across the layers, and fusing a content class into it, do not recover their cost on this corpus. Above 560 bits per frame the schemes converge. A hardware-in-the-loop 5G testbed confirms feasibility.
Low-altitude wireless networks (LAWNs) are transforming UAV swarms from distributed sensing platforms into cooperative multi-agent systems for intelligent scene understanding. However, dynamic viewpoints and partial observability often introduce ambiguity into observations, making it critical to identify the complementary semantics required to reduce the receiver agent's inference uncertainty. Existing methods typically prioritize sender-side saliency or attention responses, which may select visually prominent but redundant features while overlooking weak yet task-critical cues. To address this limitation, we propose a demand-driven primitive reasoning framework that treats the receiver's semantic deficiency as the organizing principle of multi-UAV collaboration. The ego UAV first generates an on-demand semantic request to expose its evidence gap, which guides neighboring UAVs to reorganize their local tokens into role-aware collaborative primitives. These primitives distinguish target-centric object evidence from support-centric contextual cues, forming coherent semantic units for inter-agent coordination. We further introduce a task-grounded utility model that estimates primitive contribution through counterfactual task degradation, enabling primitive selection according to their marginal value for receiver-side inference. Experiments on UAV3D and CoPerception-UAV show that our method consistently improves 3D object detection and multi-object tracking, while maintaining compact inter-agent interaction. These results indicate a shift from sender-centric feature exchange to receiver-conditioned primitive reasoning, offering a paradigm for robust multi-agent collaboration.
Radio map estimation (RME) in intelligent reflecting surface (IRS)-aided environments is challenging because the propagation field depends on both site-specific geometry and IRS reflection control. Although ray tracing can generate accurate IRS-aided radio maps, its computational cost becomes prohibitive when several IRS configurations or dense receiver grids must be evaluated. Moreover, existing learning-based approaches generally rely on IRS-aided training labels, whose generation requires expensive simulations or measurements. This study proposes a physics-guided neural network (PGNN) framework for IRS-aided RME that is trained only with IRS-free radio maps and performs zero-shot inference specifically with respect to IRS-aided received-power labels. The direct/line-of-sight (LoS), first-reflection, and diffuse-residual components serve as physics-guided representational bases for learning the aggregate received-power distribution, while propagation-oriented features are constructed from voxelized geometry. At inference time, the IRS phase- gradient configuration is converted into effective normal vectors, enabling IRS-aided radio-map estimation without IRS-aided labels or parameter retraining. Wireless InSite-based simulations show that the proposed method reproduces site-specific power variations more accurately than the stochastic and 3 rd Generation Partnership Project LoS/ non-LoS baselines.
Despite the success of deep learning in channel state information (CSI) compression and feedback, traditional models typically rely on fixed configurations, resulting in restricted generalization across practical scenarios. To address this, we propose FoundCsiNet, a wireless foundation model designed to achieve robust generalization across heterogeneous channel conditions and flexible multi-rate adaptability over multiple predefined spatial-delay dimensions and compression ratios. First, we introduce a unified tokenization module to effectively transform CSI matrices with varying spatial-delay dimensions into variable-length CSI embedding sequences, thereby guiding the model to capture shared multipath propagation characteristics across environments. To resolve the intrinsic trade-off between preserving high-fidelity details under high feedback rates and enforcing aggressive compression under low feedback rates, we develop a compression-ratio-adaptive dynamic feed-forward network (CAD-FFN) architecture. By dynamically routing representations to compression-ratio-specific parallel sub-networks, CAD-FFN combines shared multi-head self-attention with rate-specific normalization and FFN branches, thereby enabling efficient multi-rate adaptation within a unified backbone. Furthermore, we introduce a confidence-aware gated early-exit (CAGE) module to reduce complete-decoder computation adaptively based on real-time channel properties. Extensive simulations demonstrate that the proposed FoundCsiNet achieves highly robust reconstruction across heterogeneous channel environments. Remarkably, it exhibits robust out-of-distribution (OOD) generalization on held-out channel distributions without target-domain adaptation, remaining competitive with or outperforming conventional models trained on the target distributions. Further, beyond the high-accuracy reconstruction, FoundCsiNet demonstrates strong downstream versatility. By exclusively few-shot fine-tuning the decoder at the base station (BS), it aligns foundational CSI representations with downstream precoding targets, improving achievable rates while leaving the UE encoder and feedback interface unchanged.
Deep learning (DL)-based channel state information (CSI) feedback has shown strong potential for reducing feedback overhead. However, its performance often degrades in unseen propagation environments due to limited cross-environment generalization. To address this issue, we propose EnvCsiNet, an environment-aware CSI feedback foundation model that uses scene priors to adapt CSI reconstruction across propagation environments without online retraining. EnvCsiNet employs a hypernetwork-based scene-conditioned architecture to realize environment-aware adaptation by mapping scene graph representations into environment-aware reconstruction parameters that condition the CSI reconstruction process. We further introduce a lightweight scene encoding design to improve environmental processing efficiency by learning a compact representation from the high-dimensional scene graph for efficient parameter generation while preserving key structural information. Simulation results demonstrate the effectiveness of the proposed approach. EnvCsiNet achieves up to $11.7\%$ improvement in normalized mean square error over the general feedback neural network and matches the performance of the online training scheme without incurring environment-specific data collection costs.
