AI-Defined Air Interfaces in Wireless Systems

• AI-defined air interfaces are wireless PHY/MAC layers enhanced with ML techniques that dynamically adapt transmission parameters for improved communication efficiency.
• They leverage end-to-end neural network architectures, supervised and reinforcement learning, and optimization models to reduce latency, boost throughput, and enhance reliability.
• Key applications include autonomous systems, extended reality, and federated learning, achieving throughput gains of 40–60% and ultra-reliable performance metrics.

AI-defined air interfaces—also referred to as AI-native or AI-enabled air interfaces—are physical and medium access control (PHY/MAC) layers in wireless communication systems whose adaptation, control, and optimization are governed directly through ML models. By embedding data-driven intelligence at the core of modulation, coding, beamforming, and resource allocation, AI-defined air interfaces transcend the rigid, heuristic-driven architectures of 4G and 5G, enabling ultra-reliable, low-latency, and semantically rich communication critical for next-generation (NG) use cases such as extended reality, autonomous systems, and large-scale federated learning. This article details the architectural principles, mathematical frameworks, representative technologies, application scenarios, performance results, and standardization pathways that define the field.

1. Core Principles and Architectural Foundations

AI-defined air interfaces target end-to-end, learning-based optimization of the communication pipeline, encompassing signal generation, adaptive transmission, and interpretation at the receiver. Central to this paradigm are two core design principles: compression and adaptation.

• Compression refers to extracting task-relevant latent semantics from the source data, replacing bit-exact coding with information-theoretic mappings such as autoencoders or foundation models that minimize distortion with respect to high-level objectives. A formal information-bottleneck approach is used:

$$L_{\mathrm{IB}} = I(Z; X) - \beta I(Z;Y_{\text{task}})$$

where $X$ is the input, $Z$ the latent, and $Y_{\text{task}}$ the intended output or semantic target (Zhang et al., 21 Aug 2025).

• Adaptation requires that the interface dynamically reconfigures in response to channel state, user context, and application semantics. The transmission parameters $\theta$ are generated as functions of both semantic features and the channel state $h$:

$\theta = \pi(Z, h)$

allowing for scalable policy-driven adaptation across data types and link conditions (Zhang et al., 21 Aug 2025).

AI-defined air interfaces are typically realized via end-to-end neural network architectures, with modules responsible for semantic extraction, adaptive policy selection, and channel-aware mapping, often supported by a shared semantic knowledge base (SKB), and multi-modal sensory integration (Jiao et al., 15 May 2025, Zhang et al., 2024).

2. Mathematical Models and Optimization Formulations

The transition from conventional to AI-defined air interfaces is characterized by a shift from manually designed, block-wise heuristics to joint, data-driven optimization under strict physical and real-time constraints.

Generic optimization objective for AI-driven resource allocation combines network utility and cost/latency terms:

$$\max_{a, p} \quad U(\{\mathrm{SNR}_u\}) - \lambda L(a, p)$$

subject to binary activation constraints, power budgets, and URLLC requirements (Fang et al., 3 Nov 2025).

In the context of pinching antennas (PAs), system operation is modeled by per-pinch activations $a_{k,n} \in \{0,1\}$ and power allocations $p_{k,n,u}$, with user $u$’s received signal:

$$y_u = \sum_{k,n} a_{k,n} \sqrt{p_{k,n,u}} h_{k,n,u} x_u + w_u$$

where $h_{k,n,u}$ encapsulates path gain and phase, and interference is managed in the denominator of the SNR expression (Fang et al., 3 Nov 2025).

In 3GPP-oriented systems, the optimization of joint PHY/MAC variables $\mathbf{p}, \boldsymbol{\theta}, \mathbf{x}$ is formalized as:

$$\underset{\mathbf{p},\boldsymbol{\theta},\mathbf{x}}{\text{maximize}} \quad \sum_{u=1}^U B_u \log_2\bigl(1 + \mathrm{SINR}_u (\mathbf{h}, \mathbf{p}, \boldsymbol{\theta})\bigr)$$

subject to power, scheduling, latency, and inference-frequency constraints (Kontes et al., 13 Jun 2025).

For tasks such as CSI compression, minimax rate-distortion autoencoders are trained to minimize:

$$\min_{\theta, \phi} \mathbb{E}_{H \sim \mathcal{D}}\|H - g_\phi(f_\theta(H))\|_F^2$$

where $H$ is the channel matrix, $f_\theta$ the UE-side encoder, and $g_\phi$ the gNB-side decoder (Lin, 2023, Guo et al., 2022).

3. Enabling Technologies and Control Methodologies

AI model selection is dictated by task and deployment scenario:

• Supervised learning is used for tasks where oracle labels are available (e.g., beam index prediction, precoder computation).
• Reinforcement learning (RL) is fundamental for real-time control and adaptation. Markov decision process formulations specify state $s_t$, action $a_t$, and reward functions tailored to throughput, latency, or semantic accuracy (Fang et al., 3 Nov 2025).
  • Deep deterministic policy gradient (DDPG), multi-agent DDPG (MADDPG), and deep Q-networks (DQN) are commonly applied for AI-controlled air-interfaces (e.g., PA activation) with real-time updates triggered by CSI and mobility changes.
• Federated learning (FL) and over-the-air computation (AirComp) leverage the superposition nature of wireless channels for distributed ML aggregation, driving down communication rounds and latency by optimizing physical layer parameters to equalize or weight links (Fang et al., 3 Nov 2025).

Semantic communication is explicitly addressed through minimization of end-to-end semantic distortion:

$$\min_{a,E,D} \mathbb{E}_{p(x)}[d_{\text{sem}}(x, D(F(x;a), a))] + \lambda \mathrm{Cost}(a)$$

where $F,D$ are semantic encoder/decoder networks, and $a$ encodes the current air-interface configuration (Fang et al., 3 Nov 2025, Zhang et al., 21 Aug 2025).

For multi-modal and language-guided air interfaces, universal transformer backbones (e.g., LLMs adapted with LoRA/fine-tuned adapters) merge radio modality tokens and task instructions, enabling flexible output heads for positioning, classification, and beam selection tasks (Jiao et al., 15 May 2025).

4. Representative Use Cases and Application Scenarios

AI-defined air interfaces are deployed across a spectrum of functions:

• Pinching Antenna Systems (PAS): AI dynamically activates discrete dielectric pinches along a waveguide to steer LoS beams, enabling sub-wavelength spatial resolution and supporting simultaneous multi-user links. RL policies optimize activation and power jointly to satisfy URLLC and maximize sum-rate (Fang et al., 3 Nov 2025).
• CSI feedback and beam management: Autoencoders and RNN-based predictors reduce feedback by up to 60% without accuracy loss. AI-driven beam prediction and tracking significantly reduce overhead and improve mobility (Lin, 2023, Guo et al., 2022).
• End-to-End PHY Chains: Full-chain neural transceivers integrating modulation, coding, MIMO, and equalization exhibit up to 3 dB SNR gains and enable pilotless operation when jointly trained (Hoydis et al., 2020, Wang et al., 16 Mar 2025).
• Integrated Sensing and Communication: Dual-objective RL optimizes both communication rate and Cramér–Rao Bound (CRB)-based sensing accuracy, with AI-driven switches between beamforming roles (Fang et al., 3 Nov 2025).
• Multi-modal and task-oriented operation: Architectures such as AI²MMUM leverage LLMs with radio-modal adapters to handle direct positioning, LOS/NLOS classification, MIMO precoding, and beam selection via language-guided prompts and prefix tuning (Jiao et al., 15 May 2025).

5. Performance Evaluation and Benchmarking

Performance assessments of AI-defined air interfaces show:

Metric	AI-Defined Interface	Conventional Baseline
Throughput gain	+40–60%	–
Latency reduction	Up to 30% (URLLC)	–
Reliability	Outage < $10^{-5}$, $D_{\max}\! =\! 5\,\text{ms}$	–
Energy efficiency	+25% (bits/Joule)	–

Specific to FL tasks, optimized PA activation reduced communication rounds to ε-accuracy by 20%; in AirComp setups, aggregation error is halved (50% reduction), compressing convergence from $T\!=\!100$ to $T\!\approx\!60$ rounds (Fang et al., 3 Nov 2025). For path-loss prediction in WEI-6G AI² architectures, physics-based knowledge representations reduced inference times to $\sim2.3$ ms (fitting subframe requirements) and allowed pilot overhead reduction by 25% for equivalent CSI NMSE (Zhang et al., 2024).

6. Standardization and Deployment Pathways

3GPP began codifying AI/ML for air interfaces in Release 18 (TR 38.843), advancing from non-normative frameworks—focused on CSI feedback, beam management, and positioning—toward prospective normative signaling and model transfer procedures under study in Release 19 (Lin, 2023).

Key technical pathways:

• Stagewise integration: from AI-based single-module enhancements (e.g., CSI compression) through multi-task and end-to-end PHY optimization (Wang et al., 16 Mar 2025, Kontes et al., 13 Jun 2025).
• Model lifecycle management: standardized reporting of ModelID, UE capability signaling, versioning, RRC/MAC protocol extensions, and fallback procedures.
• AI-RAN functional entities: RAN nodes for training orchestration, model distribution, and distributed inference, with open interfaces for third-party algorithms (Wang et al., 16 Mar 2025).
• Regulatory concerns: Data governance, privacy-preserving FL, interoperability (two-sided model splits), safety, and trustworthiness (Kontes et al., 13 Jun 2025, Lin, 2023).

Open challenges include universal cost metrics for AI deployment (balancing capability, quality, and cost), generalization and online adaptation under non-stationary channels, and robust cross-layer optimization links with application semantics (Wang et al., 16 Mar 2025, Zhang et al., 21 Aug 2025).

7. Future Directions and Open Research Topics

Prominent avenues for future research and deployment relate to:

• Hardware-integrated AI: Field-prototyped systems (e.g., PA arrays with embedded controllers), edge-accelerated semantic encoding/decoding, and efficient on-device inference (Fang et al., 3 Nov 2025, Zhang et al., 21 Aug 2025).
• Hierarchical and meta-learning: Scaling control and resource allocation across hundreds of radio units, with adaptive policies for scenario transfer (Fang et al., 3 Nov 2025, Kontes et al., 13 Jun 2025).
• Semantic KPIs and intent alignment: Standardization beyond bit/packet error, toward semantic fidelity, task-oriented metrics, and alignment with cross-domain application goals (Zhang et al., 21 Aug 2025).
• Security and privacy: Adversarial-robust RL for jamming/spoofing resilience, federated policy learning across operators, and semantic encryption (Fang et al., 3 Nov 2025, Zhang et al., 21 Aug 2025).
• Unified simulation frameworks and digital twins: For generalized benchmarking, model stress-testing, and standardization of datasets and KPIs (Zhang et al., 21 Aug 2025, Lin, 2023).

Continued progress will demand flexible, modular architectural blueprints for hybrid AI/hardware systems, standardized performance/complexity benchmarks, and seamless integration within 3GPP and O-RAN specifications. Addressing these challenges is essential for realizing the promise of AI-defined air interfaces as the foundation of future wireless connectivity.