AI-Native 6G Networks

• AI-native 6G networks are communication systems that integrate AI as a built-in substrate, enabling continuous learning, semantic communication, and autonomous control.
• They converge diverse layers such as RAN, edge/cloud, and digital twins to co-design communication, computing, and sensing with native data collection and model lifecycle management.
• Empirical testbeds and field trials show significant improvements in latency, spectral efficiency, and security, demonstrating the practical viability of AI-native architectures.

Searching arXiv for papers on AI-native 6G networks to ground the article in current literature. I’ll check recent arXiv entries on AI-native 6G, including RAN, semantic communication, digital twins, agentic control, and security/privacy. AI-native 6G networks are communication systems in which artificial intelligence is architected as a built-in, always-on capability across radio access, core functions, edge/cloud execution, data governance, and service orchestration, rather than being attached as an external analytics module. In this literature, the term also denotes a shift from bit-centric optimization toward systems that natively expose data, semantics, reasoning, and control loops to AI models, so that communication, computing, sensing, and intelligence are co-designed from the outset (Shiwen et al., 1 Sep 2025, Wu et al., 2021, Feng et al., 4 Dec 2025, Li et al., 11 Jul 2025).

1. Conceptual foundations and defining characteristics

Early architectural work framed native AI in 6G through the notion of “intelligence inclusion”: AI services should be accessible anytime, anywhere, by anyone, and the network should treat data, computing, and algorithms as internal design primitives rather than auxiliary functions (Wu et al., 2021). Later work sharpened this position by contrasting AI-native 6G with 5G-style AI add-ons. In that contrast, traditional networks “use AI” for offline analytics, prediction, or local optimization, whereas AI-native 6G embeds AI into the network infrastructure itself, enabling continuous learning, real-time execution, autonomous control, and closed-loop adaptation across protocol layers and network domains (Shiwen et al., 1 Sep 2025, Feng et al., 4 Dec 2025, Li et al., 11 Jul 2025).

Several complementary formulations recur. One formulation presents 6G as an “endogenous intelligent” system with self-learning, self-optimization, and self-regulation capabilities, contingent on rich, real-time, structured telemetry from all protocol layers and network functions (Shiwen et al., 1 Sep 2025). Another formulation emphasizes that intelligence and meaning are integral to the communication and control fabric: AI-native RAN/edge networks should communicate features, intents, and knowledge rather than raw bits alone, and should optimize task success and semantic utility rather than only throughput, BER, BLER, or SINR (Feng et al., 4 Dec 2025). A further systems view characterizes 6G as moving from “Network for AI” to “AI for Network,” replacing siloed task-specific models with a unified foundation model and collaborative multi-agent systems for resilient and autonomous operation (Wu et al., 20 May 2026). In operator-facing standardization language, AI-native RAN is defined by three essential capabilities: AI-driven RAN processing/optimization/automation, reliable AI lifecycle management, and AI-as-a-Service provisioning (Li et al., 11 Jul 2025).

This convergence does not imply a single doctrine. Some strands focus on data-plane observability and closed-loop control, some on semantic communication and agentic intelligence, some on foundation models and multi-agent orchestration, and some on sovereign governance and security. A plausible implication is that “AI-native 6G” is best understood as an umbrella architectural program whose common denominator is the elevation of AI from optimization tool to native network substrate.

2. Architectural substrates: planes, nodes, telemetry, and digital twins

A recurring architectural baseline is the separation of connectivity, data, and intelligence into explicit system functions. One influential proposal introduces a network function plane for converged communication and computing, an independent data plane for data governance, storage, and provisioning, and an intelligent plane for Network AI Management and Orchestration (NAMO), with 6G consequently positioned as an “everything as a service” platform spanning AI, data, compute, and connectivity (Wu et al., 2021). In this view, AI workflows are first-class citizens: data collection, preprocessing, training, inference, model update, deployment, and lifecycle control are network-level functions rather than external pipelines.

More implementation-oriented work makes the telemetry substrate concrete. A real-time data-collection design for AI-native 6G uses a three-layer testbed architecture: an infrastructure layer with X86 hardware and Linux OS, an OAM layer with Docker and Kubernetes, and a network-entity layer with containerized OpenAirInterface5G RAN and Open5GS core functions, plus AI models deployed as containers in the same environment (Shiwen et al., 1 Sep 2025). Its central design choice is to move from asynchronous, external DPI-style packet capture to parallel in-NE collection: a main thread processes the bitstream while a data-collection thread reads shared buffers and writes structured JSON, thereby collecting protocol-layer and system-state data without disturbing user-plane traffic (Shiwen et al., 1 Sep 2025). The collected state is then aggregated as time-series vectors

$X_i = \{x_{i1}, x_{i2}, \ldots, x_{im}\},$

where $x_{ij}$ is the value of KPI $j$ aggregated over interval $i$ (Shiwen et al., 1 Sep 2025).

The same architectural logic appears in AI-native digital-twin work, which decomposes the twin into User Digital Twins, Infrastructure Digital Twins, and Slice Digital Twins. UDTs capture profile, behavioral, and networking information at edge nodes; IDTs represent topology, configuration, and performance of BSs, APs, routers, and edge/cloud servers; and SDTs maintain slice requirements, allocations, and QoS/QoE states at higher-level nodes (Wu et al., 2024). These twins are continuously synchronized with the physical network and host AI models for state prediction, pattern abstraction, and policy validation, often with simulator-in-the-loop DRL before actuation (Wu et al., 2024).

Day-1 standardization proposals add another native construct: an explicit AI Node in the RAN. The AI Node is edge-cloud deployed, serves multiple 6gNBs, stores RAN and UE models, performs AI training and pre-validation using RAN digital twins, provides AI inference for L3 and offloaded tasks, and exposes RAN AI capabilities to external consumers (Li et al., 11 Jul 2025). Together, these designs indicate that AI-native 6G depends not merely on stronger models, but on explicit architectural support for observability, data provenance, training loops, and distributed execution.

3. Semantic-native communication and the AI-native air interface

One major line of research argues that AI-native 6G cannot remain bit-centric. Semantic-native communication redefines the basic communication object from bit sequences to semantic representations such as latent features, intent tokens, and knowledge graphs (Feng et al., 4 Dec 2025). In the generic SemCom pipeline, a semantic encoder $f_\theta$ maps raw data $x$ to an embedding $z$,

$z = f_\theta(x),$

the receiver obtains $z'$, and a decoder $g_\phi$ reconstructs a task-level output

$x_{ij}$0

Semantic distortion is defined in embedding space as

$x_{ij}$1

and semantic-agentic RAN control is formulated by jointly optimizing agent reward and semantic distortion under bandwidth, energy, and latency constraints (Feng et al., 4 Dec 2025).

This literature organizes the semantic dimension into four abstraction levels: $x_{ij}$2 bits/symbols, $x_{ij}$3 features/latent embeddings, $x_{ij}$4 intent/task tokens, and $x_{ij}$5 knowledge/graphs (Feng et al., 4 Dec 2025). It simultaneously classifies agent autonomy from non-agentic optimization to hierarchical planner–executor structures, and maps control placement across PHY/MAC local loops, near-RT RIC, non-RT RIC, and cross-layer end-to-end orchestration (Feng et al., 4 Dec 2025). The result is a unified taxonomy in which “AI-native” increasingly denotes networks operating at $x_{ij}$6–$x_{ij}$7, with $x_{ij}$8 retained as physical substrate rather than as the sole optimization target (Feng et al., 4 Dec 2025).

At the air-interface level, this semantic turn is paired with the claim that the waveform itself should become a semantic degree of freedom. A native AI-driven air-interface proposal characterizes the design around two core characteristics: compression and adaptation (Zhang et al., 21 Aug 2025). Compression refers to semantic compression through learned encoders, NTC/NTSCC, semantic bases, and large-model embeddings, with explicit rate–distortion tradeoffs; adaptation refers to task-aware, modality-aware, and channel-aware remapping of semantic content to transmitted symbols (Zhang et al., 21 Aug 2025). In a satellite testbed over an AsiaSat 9 GEO link, semantic video communication via WAFI-VSC achieved transmission efficiency about 3× better than H.264 + LDPC at $x_{ij}$9, while traditional schemes exhibited cliff effect as SNR decreased and the semantic scheme degraded gracefully, retaining MS-SSIM $j$0 even at $j$1 dB SNR (Zhang et al., 21 Aug 2025).

A related physical-layer design, Orthogonal Semantic Sequency Division Multiplexing, pushes semantics into waveform construction via sequency-ordered Walsh bases and end-to-end learned symbol-to-waveform mappings (Hello et al., 10 Feb 2026). In AWGN, OSSDM with $j$2 achieves a target F1 score at an SNR 6 dB lower than semantic OFDM, and needs up to 10 dB less SNR than non-semantic OFDM to reach similar F1; for effective semantic spectral efficiency, the gain exceeds 400× over OFDM at order 1 and remains below 13× at order 6 (Hello et al., 10 Feb 2026). These results are task-specific rather than universal. They nevertheless suggest that an AI-native air interface is not only about control-plane intelligence, but about re-architecting the signal chain itself around semantic objectives.

4. Learning and control paradigms: federated, quantum, foundation-model, and agentic

Distributed learning is frequently treated as a foundational pillar of AI-native 6G because edge devices hold user-specific data and many target services are latency-sensitive and privacy-constrained. In the classical federated-learning workflow, device $j$3 trains a local model $j$4 on local data $j$5, and the server forms the global model via weighted averaging,

$j$6

so that raw data remain local while model parameters or gradients are exchanged (Shaon et al., 9 Sep 2025). The same paper argues that in heterogeneous and dynamic wireless settings, classical FL is limited by edge compute constraints, stragglers, intermittent connectivity, retraining latency under drift, and security/privacy vulnerabilities, and proposes Quantum Federated Learning as an extension that leverages quantum computing, quantum communication, and quantum cryptography for edge intelligence, network optimization, and secure aggregation (Shaon et al., 9 Sep 2025). In a case study with $j$7 devices, 20 sub-channels of 180 kHz each, and non-IID MNIST where each device holds samples from only one label, QAOA-based joint channel and power optimization achieved about 35% higher sum-rate and converged 40% faster than successive convex approximation, with faster convergence to higher training accuracy for QFL than FL (Shaon et al., 9 Sep 2025).

Another influential direction centers on pre-trained foundation models and cloud–edge–end collaboration. One proposal places an intent-aware pre-trained foundation model in the cloud, an expert knowledge graph and task-oriented AI toolkit in the native intelligence library, and lightweight edge/end models for execution under resource constraints (Chen et al., 2023). In that framework, the cloud interprets semantic queries and edge status, decomposes them into subtasks organized as a directed acyclic graph, and assigns classical or learned tools such as LS, ZF, DL-LS, or DL-ZF to edge or end nodes (Chen et al., 2023). A later BlueSky vision generalizes this into a 6G foundation model that is multi-modal, multi-task, and efficiently adaptable, coupled with collaborative multi-agent systems that autonomously diagnose, maintain, and recover the network (Wu et al., 20 May 2026).

The agentic turn makes this explicit. An agentic AI-native 6G proposal introduces a semantic control plane layered above deterministic 3GPP infrastructure and represents the semantic state at time $j$8 as

$j$9

where $i$0 is intent, $i$1 knowledge/telemetry, $i$2 policy constraints, $i$3 trust and identity state, $i$4 uncertainty/risk, and $i$5 multimodal and digital-twin context (Ferrag et al., 2 May 2026). The architecture then distributes bounded, policy-governed reasoning agents across device, edge, and core domains, with heterogeneous LLM deployments chosen because no single model simultaneously satisfies latency, throughput, and accuracy requirements under realistic constraints (Ferrag et al., 2 May 2026).

Alongside these high-level controllers, the literature also stresses trustworthy specialized AI for radio functions. A representative example is a CNN-based beam alignment engine for mmWave MIMO that predicts the best narrow beam from RSSI over a small set of wide sensing beams, while a Deep k-Nearest Neighbors wrapper provides confidence and credibility estimates for out-of-distribution detection (Khan et al., 23 Jan 2025). The proposed beam alignment engine reduces beam training overhead by 75% compared to exhaustive search while maintaining near-optimal performance in terms of spectral efficiency, reaches about 98.5% of the spectral efficiency of exhaustive 128-beam O-DFT search, and improves outlier detection robustness by up to 5× relative to softmax-based confidence (Khan et al., 23 Jan 2025). This line of work underscores that “AI-native” in 6G includes explainability and calibrated uncertainty, not just predictive accuracy.

5. Operational realizations, testbeds, and empirical evidence

The strongest evidence for AI-native 6G remains architectural and prototype-based rather than full-scale 6G deployment, but several papers provide concrete system results. In the Kubernetes-based OpenAirInterface5G/Open5GS testbed for real-time data collection, the proposed parallel in-NE collection architecture was compared against Wireshark-style external capture on the UPF. CPU usage was reduced by 32.7%, memory usage by 31.6%, and average data-collection latency by 78.4%, while the proposed method did not significantly increase resource consumption over the original UPF (Shiwen et al., 1 Sep 2025). The same system demonstrated an end-to-end traffic-prediction loop in which a TabNet regression model queried Prometheus every minute for KPI vectors spanning CN, RAN, and system metrics, and the study reported that UPF functions remained unaffected while data collection and AI prediction ran (Shiwen et al., 1 Sep 2025).

Large-scale operator evidence is provided by an AI-native RAN field trial built over 5000 5G-A base stations across 31 cities using a Centralized Computing Unit as an AI Node analogue (Li et al., 11 Jul 2025). For AI-enabled differentiated service assurance, average air-interface latency for short video was reduced from 43.0 ms to 32.0 ms, an improvement of 25.6%, while QR scanning latency was reduced from 18.5 ms to 14.5 ms, an improvement of 21.9%; traffic volume increased by 3–6% as applications exploited the more stable link (Li et al., 11 Jul 2025). For root-cause analysis of poor experience, AI-based RCA improved accuracy by 20% relative to rule-based methods, and user-type classification recall improved by 30% (Li et al., 11 Jul 2025). Although these results are framed as 5G-A validation, the paper presents them as evidence that the proposed Day-1 AI-native RAN architecture is practical and monetizable (Li et al., 11 Jul 2025).

Network slicing studies offer complementary evidence on AI-native control loops. In a highway scenario with two BSs, one UAV, and an autonomous-driving slice with a 100 ms delay requirement, DDPG-based resource reservation over planning windows reduced cumulative system cost by about 15% relative to a myopic optimization baseline by jointly accounting for resource reservation cost, slice reconfiguration cost, and delay violation penalty (Wu et al., 2021). In an AI-native digital-twin case study on multicast short video streaming, an adaptive UDT plus generative-AI-based 3C policy achieved about 10.3% higher average QoE than a hierarchical DRL 3C scheme at 2.2 MHz bandwidth (Wu et al., 2024). Taken together, these prototypes indicate that AI-native 6G is operationally meaningful when three ingredients are present simultaneously: native data collection, native model lifecycle control, and native interfaces between AI decisions and network actuation.

6. Governance, security, sovereignty, and open challenges

As AI becomes a native control fabric, governance and security cease to be peripheral concerns. One response is the concept of Sovereign AI, defined as the capability of a nation, operator, or organization to independently develop, govern, deploy, operate, and update AI systems in alignment with local laws, ethical values, and strategic priorities (Chetty et al., 8 Sep 2025). In this view, sovereignty extends beyond data localization to the whole AI stack: data governance, compute infrastructure, foundation models such as LLMs and LTMs, lifecycle management, and policy-aligned control in O-RAN Non-RT and Near-RT RICs (Chetty et al., 8 Sep 2025). Sovereign AI is consequently framed as a foundational pillar for secure, resilient, and ethically aligned AI-native 6G networks (Chetty et al., 8 Sep 2025).

A broader security perspective argues that AI-native 6G creates a heterogeneous threat landscape that cannot be handled through isolated, technology-specific solutions. A unified cross-layer framework divides the problem into five domains—Infrastructure Security, Network and Architectural Security, AI Security, Privacy Protection, and Security Management—and organizes countermeasures through Identity and Access Management, Security Orchestration and Automation, AI Governance and Assurance, Privacy Governance and Compliance, and Standards Harmonization as a security function (Barua et al., 1 Jul 2026). The threat space spans rogue RIS and malicious reconfiguration, JCAS sensing manipulation and unauthorized sensing, NTN trust failures, pilot contamination and beamforming manipulation in XL-MIMO, malicious xApps/rApps in O-RAN, slice isolation breaches, FL poisoning and backdoors, prompt injection and RAG poisoning for LLMs, semantic information leakage, privacy leakage from digital twins and XR telemetry, intelligent Zero Trust failures, PQC migration risks, and attestation forgery (Barua et al., 1 Jul 2026). Differential privacy, secure aggregation, homomorphic encryption, secure boot, remote attestation, strong IAM, and intelligent Zero Trust Architecture are therefore treated as native design requirements rather than add-ons (Barua et al., 1 Jul 2026).

Standardization challenges mirror the technical ones. AI-native RAN proposals argue that Day-1 6G specifications should already include AI Nodes, AI radio bearers, model-management procedures, collaborative AI computing, and AI service exposure, with 3GPP retaining architectural leadership while coordinating with ITU, ETSI, and O-RAN (Li et al., 11 Jul 2025). Semantic-agentic RAN work identifies semantic representation standardization, scalable trustworthy agent coordination, O-RAN interoperability, and energy-efficient AI deployment as open issues (Feng et al., 4 Dec 2025). Native AI-driven air-interface research adds the need for unified semantic metrics, compatibility with existing digital interfaces, and scalable validation platforms (Zhang et al., 21 Aug 2025). Quantum-FL research highlights quantum state fragility, incompatibility with classical protocol stacks, and NISQ-era hardware constraints as major barriers (Shaon et al., 9 Sep 2025).

These open issues collectively indicate that AI-native 6G is not reducible to the insertion of larger models into existing networks. It requires co-design of architectures, waveforms, data planes, model lifecycles, semantics, security, and governance. The literature therefore treats AI-native 6G less as a single technology than as a re-foundation of network design around native intelligence, native observability, and native control.