Native AI-Driven Air Interface Architecture

• Native AI-Driven Air Interface Architecture is a design paradigm that embeds AI within the radio stack to enable environment-aware sensing, inference, and adaptive control across PHY and MAC layers.
• It integrates digital twins, learned transceivers, and multi-agent reinforcement learning to optimize channel prediction, pilot reduction, energy efficiency, and real-time response in evolving 5G and 6G systems.
• The architecture emphasizes protocol visibility and lifecycle management, ensuring interoperable, fallback-capable, and safely orchestrated AI operations spanning UE, DU/BS, and centralized AI nodes.

Searching arXiv for papers on AI-native / native AI-driven air interface architectures, ChannelLM, AI-native RAN protocols, and related 6G air-interface foundation models. arxiv_search(query="AI-native 6G air interface architecture ChannelLM digital twin ChannelAgent 3GPP protocol foundation model", max_results=10) Native AI-Driven Air Interface Architecture denotes an air-interface design paradigm in which artificial intelligence is embedded into the radio stack as a first-class element of sensing, inference, control, and adaptation, rather than appended to isolated optimization blocks. In the research literature, the term covers architectures for environment-aware channel prediction, digital-twin-driven channel acquisition, learned PHY transceivers, semantic communication, protocol-visible AI operation, and model lifecycle management spanning UE, gNB, RIC, and OAM domains. The concept was articulated early as a progression from plug-in neural components to fully learned transceivers and MAC protocols, and has since been elaborated into concrete 5G NR evolution paths and 6G protocol frameworks (Hoydis et al., 2020, Lin, 2023, Lin, 25 Jun 2026).

1. Conceptual foundations and evolution

The foundational vision treats the air interface as a sequence of transmitter, channel, receiver, and MAC-layer functions that can be progressively redesigned by AI. In the early 6G formulation, Phase 1 replaces individual blocks such as the demapper or channel estimator with small neural networks; Phase 2 replaces multiple serial blocks jointly with a neural receiver; and Phase 3 jointly trains transmitter waveform and receiver in an autoencoder fashion, possibly removing hand-crafted pilot structures. The same vision extends upward to MAC, where multi-agent reinforcement learning is used for channel-access policy and signaling vocabulary formation (Hoydis et al., 2020).

The 3GPP Release-18 study on AI for 5G NR translated this idea into a lifecycle architecture with explicit modules for data collection, model training, model storage, model distribution, inference, and management and life-cycle control. That study centered the air interface on use cases such as CSI feedback enhancement, beam management, and positioning, and positioned Release 19 as the transition from study to normative specification for several AI/ML procedures (Lin, 2023).

Subsequent position work generalized the notion of nativeness beyond individual PHY algorithms. Intelligence is placed not only at higher MAC/RRC layers but also at PHY; models may be fully centralized, pushed to the DU or UE as lightweight inference kernels, or split between UE and BS; and a shared “radio-foundation” backbone may support multiple heads for beam management, CSI forecasting, and uplink power control (Kontes et al., 13 Jun 2025). This suggests that “native” describes an architectural property: sensing, inference, control semantics, retraining, and fallback are co-designed with the air interface itself.

2. Architectural strata and functional placement

A recurring structural pattern is a layered architecture in which conventional RF and protocol functions remain present, but AI capability modules are inserted at multiple granularities. One formulation distinguishes three classes of AI capability: single-module enhancers for individual PHY functions, multi-module joint optimizers that replace several serial blocks, and low-complexity solvers for mathematically difficult operations such as approximate matrix inversion or unfolded WMMSE-like procedures. Above these sits an AI management layer for data collection, training and re-training, model partition and distribution, and run-time inference scheduling (Wang et al., 16 Mar 2025).

Operator-oriented Day-1 6G proposals instantiate this abstraction as a three-tier RAN. At the edge are 6gNBs split into RU, DU, and CU; above them is an AI Node or Centralized AI Controller collocated with regional edge clouds; and above that sit central RAN OAM and the core network. Low-latency inference and data collection reside at DU/RU, while heavier training and non-real-time inference run at the AI Node. A new “AI radio bearer” on Uu carries CSI reports, raw measurement samples, and model parameters, and an Xn-like interface between 6gNBs and the AI Node carries model-management commands, telemetry, and federated-learning updates (Li et al., 11 Jul 2025).

Architectural stratum	Functions	Typical locus
AI capability modules	single-module enhancers; multi-module joint optimizers; low-complexity solvers	UE, RU, DU, BS PHY
AI management layer	data collection; training/re-training; model distribution; inference scheduling	AI Node, edge cloud, OAM
RAN execution plane	beamforming; scheduling; channel estimation; interference control	RU, DU, CU

This placement is not merely organizational. It is tied to latency classes: sub-millisecond or slot-scale inference is localized, whereas cross-cell optimization, large-model adaptation, and long-horizon analytics are centralized. A plausible implication is that native AI architectures are defined as much by task placement and lifecycle partitioning as by the specific neural architecture used within any one block.

3. Environment-native channel intelligence

One major branch of native AI-driven air-interface research shifts channel acquisition away from purely statistical modeling toward explicit environment modeling. In the WEI-6G AI$^{2}$ framework, wireless environment information is organized into four progressive steps: raw sensing data, features obtained by dimensionality reduction, semantics tailored to tasks, and knowledge that quantifies environmental impact on the channel. The paper argues that these four steps support scenario adaptability, real-time inference, and proactive action, including reduced dependence on pilots and beam sweeping. In the path-loss case study, raw sensing required 430 ms of inference with MSE $=9.87\,\mathrm{dB}^2$, whereas environment knowledge required 2.3 ms with MSE $=5.65\,\mathrm{dB}^2$; in a CSI use case, WEI enabled the system to maintain a target NMSE using 25% fewer pilots (Zhang et al., 2024).

The ChannelLM-driven Digital Twin Channel architecture advances this line of work by making the digital twin itself the channel-acquisition engine. It consists of three modules: environment reconstruction, environment feature extraction, and a ChannelLM core. Environment reconstruction combines an offline static-scene model with online dynamic-object detection using YOLOv11 on RGB images, depth-based coarse localization, and LiDAR/DBSCAN-based fine 3D cuboid localization. Feature extraction produces BS/UT location maps, scatterer distribution and height maps, and a penetration-ratio map. Representative definitions are

$B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$

$$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$

and, for blocked LoS,

$PR(q)=\frac{\|q_i-q_o\|}{\|q_i-q\|}.$

These features are processed by a GPT-2 small backbone with 12 layers, 12 heads, and hidden dimension 768, followed by separate PL and CSI heads (Cai et al., 20 Apr 2026).

The reported performance of ChannelLM is explicitly environment-generalizable. In unseen static scenarios, PL-map RMSE is $4.44\,\mathrm{dB}$, compared with $5.87\,\mathrm{dB}$ for PMNet and $6.31\,\mathrm{dB}$ without GPT-2. For CSI prediction in unseen scenarios, a small model without environment features gives $\mathrm{NMSE}\approx-8.91\,\mathrm{dB}$ and $=9.87\,\mathrm{dB}^2$0, whereas ChannelLM gives $=9.87\,\mathrm{dB}^2$1 and $=9.87\,\mathrm{dB}^2$2, corresponding to a $=9.87\,\mathrm{dB}^2$3 reduction in prediction error. In dynamic operation, YOLOv11 detection, depth localization, LiDAR clustering, feature extraction, and ChannelLM inference sum to approximately 70 ms end-to-end (Cai et al., 20 Apr 2026).

An agentic extension appears in the ChannelAgent-empowered electromagnetic space world model, which organizes wireless intelligence into a closed loop of multi-modal sensing, ChannelAgent as the intelligent core, and execution with feedback update. In the path-loss case study, the agent selects compact feature subsets from ten candidate features using a hybrid MDP, reinforcement-learning-inspired policy update, and evolutionary search. The task score is

$=9.87\,\mathrm{dB}^2$4

For Task 1, agent-driven selection with features $=9.87\,\mathrm{dB}^2$5 achieves RMSE $=9.87\,\mathrm{dB}^2$6, versus $=9.87\,\mathrm{dB}^2$7 for the full-feature baseline; for Task 2, features $=9.87\,\mathrm{dB}^2$8 achieve RMSE $=9.87\,\mathrm{dB}^2$9, versus $=5.65\,\mathrm{dB}^2$0 for the full-feature baseline; and in the multi-scenario task, the selected subset $=5.65\,\mathrm{dB}^2$1 yields RMSE $=5.65\,\mathrm{dB}^2$2 (Li et al., 14 May 2026).

4. Learned transceivers, foundation models, and semantic adaptation

A second branch of the field treats the air interface itself as an end-to-end learned object. One example is the adaptive pilot-free and CP-free transceiver in which the transmitter is a learnable constellation shaper and the receiver is a residual CNN with a lightweight Channel Adapter inserted after each block. The processing chain is binary bitstream $=5.65\,\mathrm{dB}^2$3 $=5.65\,\mathrm{dB}^2$4 AI-driven constellation shaper $=5.65\,\mathrm{dB}^2$5 IFFT without CP $=5.65\,\mathrm{dB}^2$6 time-varying channel $=5.65\,\mathrm{dB}^2$7 FFT $=5.65\,\mathrm{dB}^2$8 neural receiver $=5.65\,\mathrm{dB}^2$9 bit-wise LLRs. Training uses a composite cross-entropy loss plus a PAPR penalty under an augmented Lagrangian. The architecture is reported to achieve a $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$0 gain at $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$1 over prior pilot-free CP-free designs in the CDL-C model at 120 km/h, 26.4% higher throughput than the 5G baseline, near-full adaptation with only 3.5% of full-model parameters fine-tuned on new channels, and $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$2 BER loss when constraining $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$3 for PAPR compliance (Cheng et al., 29 Oct 2025).

A foundation-model formulation appears in AirFM-DDA, which reparameterizes CSI from the space-time-frequency domain into the delay-doppler-angle domain using a four-dimensional symplectic Fourier transform and 2D FFT over the receive array. The purpose is to resolve multipath explicitly along physically meaningful axes before applying Transformer-based learning. The model adds frame-structure-aware positional encoding and replaces global attention with window-based attention, whose complexity scales as $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$4 rather than $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$5. Empirically, AirFM-DDA is reported to achieve superior zero-shot generalization on prediction and estimation tasks across unseen scenarios and datasets, while window-based attention delivers 8× faster inference, 31 ms versus 246 ms, and 5× lower GPU memory, 2.0 GiB versus 11.1 GiB, relative to global attention (Bian et al., 19 Apr 2026).

Native AI-driven architecture is also formulated in semantic terms. A representative design uses an end-to-end trainable transmitter-receiver pair augmented by a semantic knowledge base and a transmissive adaptation engine. The transmitter semantic encoder maps raw source data $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$6 to a low-dimensional semantic code $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$7; the adaptation engine uses real-time CSI, task context, and data-type metadata from the semantic knowledge base to steer modulation order, resource allocation, power allocation, and encoder selection; and the receiver performs the inverse adaptation and semantic decoding. The paper defines the two core characteristics as compression and adaptation and validates them on a GEO satellite link using AsiaSat 9. In that testbed, semantic video transmission yields 3× higher efficiency than H.264+LDPC at $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$8; above $B_j(q)=1 \text{ if } q=q_j,\ \text{else } 0,$9, all methods tie in MS-SSIM versus SNR; below $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$0, H.264/H.265 exhibit the cliff effect while semantic coding degrades gracefully to $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$1 at $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$2, and remains visually intelligible at $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$3 (Zhang et al., 21 Aug 2025).

Taken together, these lines of work show that native AI at the air interface does not refer to one single design idiom. It includes channel-centric digital twins, physically structured foundation models, pilot-free end-to-end transceivers, and semantic transmitter-receiver co-design.

5. Protocol visibility, lifecycle management, and control-plane realization

A central architectural issue is how AI behavior becomes interoperable across vendors and deployments. The 3GPP-oriented protocol framework argues that standardization should not prescribe model architectures, training methods, or model weights. Instead, it should define protocol semantics for AI-enabled operation: AI Task Profiles, Capability Descriptors, Profile Configuration, MAC Control, PHY Signaling, and OAM/M-Plane interfaces. The corresponding RRC AIProfileConfig information element carries ProfileID, TaskID, InputDescriptor, OutputDescriptor, Timing, ValidityRegion, FallbackProfileID, and Version. MAC control elements support ActivateProfile, SuspendProfile, SwitchProfile, and FallbackCommand; UCI carries ProfileReady, ValidityRequest, or ConfidenceIndex; and every AI profile is tied to a conservative anchor through an explicit fallback profile (Lin, 25 Jun 2026).

This protocol-visible view extends the Release-18 5G NR AI/ML framework, which already treated data collection, training, storage, distribution, inference, and life-cycle management as air-interface functions. That study proposed control-plane signaling for model IDs, activation and switching, and described a standardized 16-bit ModelID together with an “AI-ModelConfig” information element carrying $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$4. It also separated control-plane model management from data-plane inference, with monitoring metrics and fallback to legacy procedures when AI performance degrades below threshold (Lin, 2023).

A more radical demonstration replaces rule-based RRC state-machine behavior with a large AI model in the CU-CP. In that system, ASN.1 RRC messages are treated as a domain-specific language; messages are linearized, byte-pair encoded, and processed by a decoder-only LLaMA-class model fine-tuned with LoRA. On $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$5 real-world request-response pairs, the resulting 8B model achieves median cosine similarity $$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$6 relative to ground-truth field messages, a 61% relative gain over a zero-shot LLaMA-3 8B baseline. However, median inference latency is 6.9 s per message, average latency is 10.4 s, and VRAM use is approximately 29 GB per thread on an edge A100 GPU (Liu et al., 22 May 2025). This suggests that protocol literacy can be learned from operational traces, while real-time protocol execution at control-plane timescales still requires substantial compression, specialization, or architectural restructuring.

6. Optimization criteria, standardization trajectory, and unresolved issues

As native AI spreads across the air interface, several works shift from isolated accuracy metrics to system-level design criteria. One proposed criterion balances AI-enabled communication capability, AI model quality, and AI-induced cost:

$$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$7

A multi-scenario extension weights capability, quality, and cost across deployment environments, and the paper recommends a two-step solution: first impose hard constraints on quality and delay, then optimize the multi-scenario criterion over the feasible set. The same work organizes standardization into three phases: single-module use cases, multi-module benchmarks, and full-chain AI from PHY to MAC and scheduling (Wang et al., 16 Mar 2025).

Large-scale operator results indicate that these architectural ideas are already being evaluated at scale. In a nationwide trial over more than 5000 5G-A base stations, average air-interface latency fell by 25.6% for short-video streaming, from 43 ms to 32 ms, and by 21.9% for QR scanning, from 18.5 ms to 14.5 ms. Energy saving reached 34.16% when cell-level traffic prediction was combined with service-aware aggregation, and root-cause classification accuracy improved from 70% to 90% with recall increased by 30%. The average DU inference overhead per TTI was under 0.5 ms and consumed 15% of a 4-core CPU, while AI Node GPU utilization rose by 5% (Li et al., 11 Jul 2025).

Another emerging direction is agentic orchestration of AI-native RAN functions through semantic intents. In the energy-efficient AI-RAN architecture, a semantic intent layer, an LLM-driven semantic coordinator agent, a digital twin agent, and a configuration-deployment-monitoring agent jointly map operator objectives into validated RAN actions. The framework models total energy as

$$H(q)=h_{\max}(q) \text{ if } q \text{ lies inside a scatterer, else } 0,$$8

and combines performance, latency, and energy in a multi-objective cost. In the reported use cases, predictive sector steering achieved 19.61% energy gain versus an All-ON baseline while maintaining throughput satisfaction at 95.82% and coverage satisfaction at 100%; DRL-based carrier steering yielded 15% and 18% energy saving under UX-95 and UX-90 profiles, respectively, with coverage at or above 99% (Aroua et al., 20 Jun 2026).

The open issues identified across the literature are consistent. WEI-based work highlights multi-modal data synchronization, knowledge construction for diverse tasks, and integration with legacy protocol stacks as unresolved problems (Zhang et al., 2024). 3GPP-oriented protocol work argues that a common misconception is to equate AI-native standardization with standardization of model internals; its counterposition is that the standard should define task semantics, timing, validity, confidence, activation, monitoring, and safe fallback, while preserving implementation freedom (Lin, 25 Jun 2026). Another recurrent misconception is that nativeness requires wholesale replacement of legacy procedures. Published frameworks instead retain fallback profiles, observation or shadow modes, and conventional anchors, indicating that native AI is being pursued as an interoperable and safely revertible air-interface architecture rather than as an all-or-nothing departure from protocol history.