Large AI Models (LAM)

• Large AI Models (LAM) are foundation-scale neural architectures with billions to trillions of parameters, trained on massive multimodal datasets.
• They enable superior transfer learning, zero-shot generalization, and emergent cognitive capabilities across diverse applications.
• LAMs enhance wireless, IoT, and semantic communications by improving channel estimation, beamforming, and edge computing efficiency.

Large AI Model (LAM) refers to foundation-scale neural architectures—typically with billions to trillions of parameters—trained on massive, often multimodal datasets, and exhibiting emergent cognitive and generalization abilities. LAMs are distinguished from conventional AI models by extreme parameter scale ($10^9$–$10^{12}$), superior transfer and zero-shot capabilities, and robust fusion of data modalities such as text, vision, audio, and sensor inputs. LAMs are increasingly deployed as central components in wireless physical layer systems, semantic communications, resource management, edge computing, IoT, and network automation, marking a fundamental shift in the intelligence, adaptability, and operational efficiency of next-generation (6G/NTN) communication networks (Guo et al., 4 Aug 2025, Li et al., 17 Dec 2025, Jiang et al., 6 May 2025).

1. Foundational Definition and Core Differentiators

A Large AI Model (LAM) is formally characterized by its parameter scale—billions to trillions—vast computational budgets (often $10^3$–$10^5$ GPU-years), and training on diversified multi-domain datasets (Guo et al., 4 Aug 2025, Jiang et al., 6 May 2025). LAMs manifest:

• Generalization: Zero/few-shot task adaptation and resistance to distribution shifts, overcoming poor generalization in small models.
• Multimodality: Native processing for heterogeneous inputs (e.g., CSI, images, lidar, text) within unified architectures (Transformers, Diffusion, State-Space Models).
• Scalability: Orders-of-magnitude increases in model and dataset size correlate with power-law improvements in downstream task metrics (MSE, perplexity).
• Emergent capabilities: Chain-of-thought reasoning, tool invocation, and agentic interactions without explicit coding or scripting (Jiang et al., 28 May 2025).

The distinction between LAMs and conventional AI models is further summarized:

Property	Conventional AI Models	Large AI Model (LAM)
Parameter scale	$10^{5}$–$10^{7}$	$10^{9}$–$10^{12}$
Data modality	Typically unimodal	Multimodal fusion
Generalization	Task-specific, fragile	Universal, robust
Adaptation	Extensive retraining	Few-shot / prompt / plug-in
Deployment	Monolithic	Distributed, microservice, edge

2. Architectures and Mathematical Frameworks

LAMs employ deep stacks of Transformer encoder/decoder blocks, often interleaved with convolutional or state-space modules to capture local structure. Key mathematical formulations include:

• Transfer-learning LAMs: Wireless data mapped to embedding/token spaces, optionally fine-tuned via

$$L_{\mathrm{FT}}(\theta) = L_{\mathrm{task}}\bigl(\theta; \mathcal{D}\bigr) + \lambda_{\mathrm{reg}}\,\lVert \theta - \theta_{0}\rVert_{2}^{2}$$

for task-specific or regularization objectives (Guo et al., 4 Aug 2025).

• Native wireless LAMs: Pretraining via masked signal modeling, using loss functions of form

$$\mathcal{L}(\theta) = \alpha\,\mathcal{L}_{\mathrm{comm}}(\theta) + \beta\,\mathcal{L}_{\mathrm{multimodal}}(\theta) + \gamma\,\mathcal{L}_{\mathrm{reg}}(\theta)$$

where each component aligns with communication efficacy, multimodal coherence, and regularization (Guo et al., 4 Aug 2025).

• Parameter scaling laws: Empirically,

$N \propto C^{\,0.75}, \qquad C \propto N^{\,1.3}$

relating model size $N$ and compute $C$ to observed improvements in prediction metrics.

• Adaptive compression: Semantic distortion is minimized given model size constraints:

$$\min_{\theta'} D_S(\theta,\theta') + \lambda C(\theta') \quad \text{s.t.}\quad C(\theta') \leq R$$

enabling deployment in resource-limited and edge environments (Ni et al., 28 Mar 2025, Lyu et al., 14 May 2025).

3. Application Strategies and Representative Performance

LAM application to physical layer communications is delineated into two principal strategies (Guo et al., 4 Aug 2025):

1. Leveraging pre-trained LAMs: Utilizing models such as GPT-4 or GPT-2 backbone for classification, denoising, or regression via transfer-learning, often with frozen parameters and light task heads.
2. Developing native LAMs: Architectures built from scratch on wireless datasets, optimized for diverse downstream tasks via multi-task heads (channel estimation, beam selection, signal reconstruction).

Representative use cases demonstrate substantial gains:

• Channel Estimation/Prediction: Pre-trained LLM4CP reduces normalized MSE from –16 dB to –22 dB on scenario changes (Guo et al., 4 Aug 2025).
• Beamforming/Precoding: M²BeamLLM achieves spectral efficiency gains of 15% and beam selection accuracy from 72% to 88% compared to codebook search (Guo et al., 4 Aug 2025).
• Modulation/Sensor Classification: WirelessGPT increases classification accuracy from 93% to 98% at SNR=10 dB, with a 40% BER reduction in fading channels (Guo et al., 4 Aug 2025).
• 3D Radio Map Estimation: RadioLAM outperforms 3D-UNet and interpolation by 40–90% in MSE at ultra-low (0.1%) sampling rates through augmentation, MoE generation, and physics-guided election (Liu et al., 15 Sep 2025).

4. Edge, Satellite, and IoT Deployment Paradigms

Deploying LAMs in resource-constrained edge, IoT, and space environments necessitates decomposition, distributed training, and microservice orchestration (Wang et al., 1 May 2025, Wang et al., 6 May 2025, Shi et al., 2 Apr 2025):

• Federated Fine-Tuning: LoRA adapters and knowledge distillation allow parameter-efficient personalization on edge/satellite devices without full model transfer.
• Microservice Inference: LAM modules (encoders, decoders, experts) virtualized as stateless services, supporting DAG orchestration for multi-task latency reduction.
• Collaborative Large–Small Model Frameworks: LASCO/E-LASCO treat the LAM as a universal prior and SAMs as plug-ins for environment-specific adaptation, enabling rapid convergence and efficient inference in multi-user air interfaces (Cui et al., 13 Dec 2025).

Latency, memory, power, and communication constraints are mitigated by optimal parameter splits, adaptive pruning, split-layer scheduling, quantization (INT4/8), and RL-driven orchestration.

5. Semantic Communications and Multimodal Alignment

LAMs represent foundational advances in semantic communication systems, facilitating content-adaptive transmission, multimodal fusion, and knowledge-centric compression (Jiang et al., 2023, Jiang et al., 2023, Ni et al., 28 Mar 2025):

• Prompt-enabled LAMs: CSI feedback models leveraging statistical distributions as prompts (mean magnitude in angular-delay domain) to enhance generalization and sample efficiency (Guo et al., 18 Jan 2025).
• Multimodal Semantic Pipelines: MMA (MLM-based alignment), LKB (LLM-guided personalization), and CGE (GAN-based channel estimation) enable loss-resilient, low-bitrate, context-consistent communication (Jiang et al., 2023).
• Knowledge Base Construction: Universal models (e.g., SAM) segment images into semantic regions for attention-based fusion and adaptive semantic compression, addressing the limitations of conventional KBs on representation and update frequencies (Jiang et al., 2023).

LAM-based semantic frameworks demonstrate drastic bit-rate savings, improved semantic fidelity, and direct handling of data heterogeneity and ambiguity, supported by multi-modal retrieval-augmented generation (Ni et al., 28 Mar 2025).

6. Agentic Systems, Protocol Automation, and Network Optimization

Agentic AI systems derived from LAMs—comprising planners, knowledge bases, tools, and memory modules—enable autonomous network orchestration, protocol emulation, and resource allocation (Jiang et al., 28 May 2025, Liu et al., 22 May 2025, Ibrahim et al., 13 Jan 2026):

• AI-native protocol generation: LAMs (e.g., LLaMA3-8B with LoRA) directly generate ASN.1-encoded RRC control messages for 6G RANs, attaining median cosine similarity of 0.97 to ground truth, a 61% gain over zero-shot baseline (Liu et al., 22 May 2025).
• Agentic orchestration frameworks: Systems integrate multi-agent collaborative planning, evaluation, and data retrieval over vector and graph KBs, enabling reflection, tool invocation, and continual improvement loops (Jiang et al., 28 May 2025).
• Resource allocation in NTN: LAM-coordinated DRL (TD3) agents outperform classical baselines by 40–64% in throughput, fairness, and outage via curriculum-style LLM-guided reward shaping and attention modulation (Ibrahim et al., 13 Jan 2026).

These frameworks support real-time adaptation, interpretability, and reproducibility across operational domains including physical-layer modulation, semantic transmission, network slicing, and security.

7. Challenges, Interpretability, and Future Research Directions

Outstanding challenges for LAMs include data scarcity, efficient environment adaptation, model explainability, real-time edge deployment, and security (Guo et al., 4 Aug 2025, Jiang et al., 6 May 2025, Wang et al., 6 May 2025, Li et al., 17 Dec 2025):

• Efficient architectures: Mixture-of-Experts, state-space transformers, aggressive pruning/quantization.
• Interpretability: Attention map analysis, SHAP/LIME feature attribution, and counterfactual decision auditing.
• Deployment and standardization: Federated/split learning protocols, standardized open-datasets, composite evaluation metrics (e.g., IBSA-Score (Li et al., 17 Dec 2025)).
• Security and governance: Privacy via differential privacy, homomorphic encryption, adversarial training, and human-in-the-loop overrides.
• Theoretical frontiers: Physics-informed LAMs, advanced reasoning modules, simulation-to-reality transfer, constrained generative models, and joint optimization across network layers.

Emerging research directions involve physics-aware regularization, lifelong continual adaptation under non-i.i.d. distributions, domain-randomized benchmarks, and integration of neuromorphic computing and generative models for network optimization.

---

References

• "Large AI Models for Wireless Physical Layer" (Guo et al., 4 Aug 2025)
• "Large Model Enabled Embodied Intelligence for 6G Integrated Perception, Communication, and Computation Network" (Li et al., 17 Dec 2025)
• "LLM-Based Emulation of the Radio Resource Control Layer: Towards AI-Native RAN Protocols" (Liu et al., 22 May 2025)
• "Large and Small Model Collaboration for Air Interface" (Cui et al., 13 Dec 2025)
• "Prompt-Enabled Large AI Models for CSI Feedback" (Guo et al., 18 Jan 2025)
• "RadioLAM: A Large AI Model for Fine-Grained 3D Radio Map Estimation" (Liu et al., 15 Sep 2025)
• "Multi-Task Semantic Communications via Large Models" (Ni et al., 28 Mar 2025)
• "Large AI Model-Based Semantic Communications" (Jiang et al., 2023)
• "Large AI Model Empowered Multimodal Semantic Communications" (Jiang et al., 2023)
• "The Larger the Merrier? Efficient Large AI Model Inference in Wireless Edge Networks" (Lyu et al., 14 May 2025)
• "Edge Large AI Models: Revolutionizing 6G Networks" (Wang et al., 1 May 2025)
• "A Comprehensive Survey of Large AI Models for Future Communications: Foundations, Applications and Challenges" (Jiang et al., 6 May 2025)
• "Large Artificial Intelligence Model Guided Deep Reinforcement Learning for Resource Allocation in Non Terrestrial Networks" (Ibrahim et al., 13 Jan 2026)
• "Edge Large AI Models: Collaborative Deployment and IoT Applications" (Wang et al., 6 May 2025)
• "Satellite Edge Artificial Intelligence with Large Models: Architectures and Technologies" (Shi et al., 2 Apr 2025)
• "From Large AI Models to Agentic AI: A Tutorial on Future Intelligent Communications" (Jiang et al., 28 May 2025)