PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey of Large AI Models for Future Communications: Foundations, Applications and Challenges

Feibo Jiang, Cunhua Pan, Li Dong, Kezhi Wang, Merouane Debbah, Dusit Niyato, Zhu Han

arXiv:2505.03556v1cs.ITmath.IT

TL;DR

Communication systems are growing more complex as 6G pursues ubiquitous, intelligent services. This survey reviews LAM foundations and communication applications, and synthesizes current research challenges and future directions.

Abstract

from arXiv · show

The 6G wireless communications aim to establish an intelligent world of ubiquitous connectivity, providing an unprecedented communication experience. Large artificial intelligence models (LAMs) are characterized by significantly larger scales (e.g., billions or trillions of parameters) compared to typical artificial intelligence (AI) models. LAMs exhibit outstanding cognitive abilities, including strong generalization capabilities for fine-tuning to downstream tasks, and emergent capabilities to handle tasks unseen during training. Therefore, LAMs efficiently provide AI services for diverse communication applications, making them crucial tools for addressing complex challenges in future wireless communication systems. This study provides a comprehensive review of the foundations, applications, and challenges of LAMs in communication. First, we introduce the current state of AI-based communication systems, emphasizing the motivation behind integrating LAMs into communications and summarizing the key contributions. We then present an overview of the essential concepts of LAMs in communication. This includes an introduction to the main architectures of LAMs, such as transformer, diffusion models, and mamba. We also explore the classification of LAMs, including large language models (LLMs), large vision models (LVMs), large multimodal models (LMMs), and world models, and examine their potential applications in communication. Additionally, we cover the training methods and evaluation techniques for LAMs in communication systems. Lastly, we introduce optimization strategies such as chain of thought (CoT), retrieval augmented generation (RAG), and agentic systems. Following this, we discuss the research advancements of LAMs across various communication scenarios. Finally, we analyze the challenges in the current research and provide insights into potential future research directions.

I. INTRODUCTION

As communication systems grow more complex, 6G’s vision of ubiquitous intelligence motivates integrating LAMs into communication services. The introduction places LAMs in the progression of AI and distinguishes their scale, capabilities, and adaptation from other models.

1) Deep learning-assisted communication:

  • Deep learning improves communication performance and efficiency, but its generalization remains limited in dynamic, uncertain environments.Communication systems still face adaptive optimization and learning challenges.Source: 1) Deep learning-assisted communication · p. 2

2) Reinforcement learning-assisted communication:

3) Generative AI-assisted communication:

1) Definition:

  • LAMs include LLMs, LVMs, LMMs, and world models, with parameter scales reaching hundreds of billions or trillions.Source: 1) Definition · p. 2

2) Distinction between LAMs and GAIs:

  • Compared with other generative AI models, LAMs are typically larger, generalize across broader tasks, and can exhibit in-context learning, CoT, and reflection.These capabilities can support downstream adaptation without task-specific retraining.Source: 2) Distinction between LAMs and GAIs · p. 2

3) Distinction between LAMs and pretrained FMs:

  • Pretrained foundation models may hallucinate and need task adaptation; domain-specific optimization can help LAMs handle communication tasks and mitigate those issues.Language-based models may require instruction fine-tuning and RLHF.Source: 3) Distinction between LAMs and… · p. 2
  • LAMs can capture global spatiotemporal dependencies and support traffic prediction, resource allocation, and decisions across heterogeneous devices and tasks.The cited examples include learning mobility and traffic fluctuations, and prompt-based offloading without retraining.Source: 3) Distinction between LAMs and… · p. 3
  • LAMs’ contextual learning and emergent reasoning can support zero- or few-shot analysis of user needs and dynamic planning for communication networks.Source: 3) Distinction between LAMs and… · p. 3

C. Related survey work

The survey identifies gaps in prior work on model coverage and communication applications, then presents a framework spanning LAM foundations, applications, challenges, and future directions.

1) Limited model coverage:

  • Prior surveys often focus on LLMs, leaving LVMs, LMMs, and world models—and their architectures, training, and alignment—undercovered.Source: 1) Limited model coverage · p. 3

2) Incomplete application landscape:

  • Prior communication surveys give limited coverage to broader scenarios and lack systematic comparisons of model suitability, technical characteristics, and collaboration strategies.Source: 2) Incomplete application landscape · p. 3

D. Contributions

  • The paper organizes its review around LAM fundamentals, applications, challenges, and future directions.Source: D. Contributions · p. 3

1) Foundations of LAMs for communications:

  • Its foundations coverage includes transformer, diffusion, and mamba architectures; LAM categories; training and alignment; evaluation; and optimization techniques.The evaluation topics include communication Q&A, tool learning, modeling, and code design; optimization includes CoT, RAG, and agentic systems.Source: 1) Foundations of LAMs for… · p. 3

2) Applications of LAMs for communications:

  • The application review covers physical-layer design, resource allocation, network management, edge intelligence, semantic communication, agentic systems, and emerging applications.It classifies research progress and discusses representative works and application prospects.Source: 2) Applications of LAMs for… · pp. 3–4

3) Research challenges of LAMs for communications:

The section reviews LAM foundations and transformer architecture while identifying data, knowledge, reliability, adaptability, latency, privacy, and security challenges in communication settings.

3) Research challenges of LAMs for communications:

  • LAMs face communication challenges spanning scarce high-quality labeled data, domain-knowledge integration, hallucinations, reasoning, interpretability, dynamic-task adaptation, inference latency, privacy, and security.The survey also links data scarcity, privacy, and cost to constraints on training and generalization.Source: 3) Research challenges of LAMs for… · p. 4

II. FOUNDATIONS OF LAMS FOR COMMUNICATIONS

  • Large-scale pretraining gives LAMs multi-task generalization, while in-context learning and fine-tuning support downstream adaptation and emergent abilities address unseen tasks.The survey frames these abilities as advantageous for future intelligent communication systems.Source: §II · p. 4

A. Key architecture of LAMs

  • The foundations section introduces transformer, diffusion, and Mamba as key LAM architectures.The supplied passage identifies these architectures but does not elaborate on diffusion or Mamba here.Source: A. Key architecture of LAMs · pp. 4, 6

1) Transformer model:

  • Transformers rely on attention for parallel sequence processing and long-range dependencies; their communication applications include large-scale MIMO, semantic communication, and wireless image transmission.The survey describes encoder self-attention with queries, keys, and values, plus decoder cross-attention; it reports a ViT semantic system outperforming traditional 256-QAM in low-SNR environments.Source: 1) Transformer model · p. 6

2) Diffusion model:

The section describes diffusion and Mamba architectures and their communication applications, then presents a data-type classification of LAMs and reviews LLM capabilities and uses.

2) Diffusion model:

  • Diffusion models add noise to data and learn reverse denoising, with communication applications spanning channel estimation, image transmission, MIMO semantic communication, RF generation, and audio semantic communication.The survey also discusses their potential in network optimization scenarios.Source: 2) Diffusion model · p. 7

3) Mamba model:

  • Mamba uses input-dependent selection and hardware-aware computation to process long sequences; its described workflow projects inputs, updates state-space parameters, computes outputs, and generates task-specific results.The survey presents Mamba as suited to high-dimensional long-series data.Source: 3) Mamba model · p. 7
  • Communication applications of Mamba include image transmission, traffic-flow estimation, and underwater acoustic image semantic communication.The survey reports MambaJSCC as achieving high performance and low latency without increased computation or parameter overhead, and ST-Mamba as performing well with limited data.Source: 3) Mamba model · pp. 7–8

B. Classification of LAMs

  • The survey classifies LAMs by the type of data processed, presenting this as a practical framework for communication tasks involving different modalities and resource-allocation needs.It says the classification can address diverse communication challenges and improve system efficiency across tasks.Source: B. Classification of LAMs · p. 8

1) Large language model:

  • LLMs learn language structure and semantics from large text corpora and support language generation, translation, question answering, memory, and multi-step reasoning.The survey describes their language capabilities as arising from pretraining on large-scale text.Source: 1) Large language model · p. 8
  • Communication uses of LLMs include multimodal semantic communication, semantic understanding and reconstruction, multi-agent network optimization, end-to-end semantic communication, and split-learning assistants.The surveyed examples include fine-tuning for computational complexity and edge offloading to address mobile-device capacity limits.Source: 1) Large language model · p. 9

2) Large vision model:

The section introduces LVM capabilities and examples, including segmentation, feature learning, and multiple visual tasks, and describes their use in communication systems.

2) Large vision model:

  • LVMs process visual data and learn image features, supporting image classification, object detection, segmentation, and generation.The survey says LVMs typically use CNN or transformer architectures and can generalize across visual tasks.Source: 2) Large vision model · p. 9
  • LVM feature learning combines CNN-extracted local features with transformer self-attention for long-range dependencies and contextual information.Large-scale pretraining and fine-tuning are described as ways to optimize feature representations for visual tasks.Source: 2) Large vision model · p. 9
  • The survey lists image recognition, detection, scene parsing, segmentation, generation, editing, video analysis, and 3D reconstruction among LVM-supported tasks.Source: 2) Large vision model · p. 9
  • The cited LVM examples are the SAM, DINO, and Stable Diffusion series.Source: 2) Large vision model · p. 9
  • SAM targets image segmentation; SAM-2 is described as improving accuracy on complex scenes and small targets and increasing inference speed.The survey also notes SAM-2's upgraded architecture and optimized self-attention.Source: 2) Large vision model · p. 10
  • DINO learns general visual features from curated datasets without fine-tuning, while DINO V2 is described as improving feature accuracy, robustness, and generalization.DINO V1 uses transformer architecture and contrastive learning on images from different perspectives.Source: 2) Large vision model · p. 10
  • Stable Diffusion models support image generation, restoration, and transformation; version 3 replaces version 2's U-Net backbone with a rectified Fourier transformer architecture.The survey says version 3 uses text and image encoding tracks to improve multimodal image interaction.Source: 2) Large vision model · p. 10
  • Communication applications include SAM-based semantic segmentation for image transmission and a SAM-based method intended to retain semantic features and support high-quality reconstruction.The LAM-SC framework uses a semantic knowledge base; the other method targets diminishing returns from raising transmission rates while reducing communication overhead.Source: 2) Large vision model · pp. 10–11

3) Large multimodal model:

The section surveys multimodal models and world models, then outlines the three-stage training process for LAMs in communications. It describes how these model families represent or simulate complex data and environments.

3) Large multimodal model:

  • LMMs process multiple modalities through cross-modal fusion and shared representations, supporting multimodal understanding, reasoning, and prediction.They may combine image, text, speech, audio, and video features in a unified representation.Source: 3) Large multimodal model · p. 11
  • Examples include CoDi’s conditional diffusion for image generation, Meta-transformer’s shared encoder across 12 modalities, and ImageBind’s shared embedding space for six modalities.ImageBind aligns modalities without explicit paired data.Source: 3) Large multimodal model · p. 11
  • Communication applications include LAM-MSC for cross-modal data processing and a delay-aware semantic framework that decomposes and transmits modalities using different schemes.The latter uses retransmission for text prompts and adaptive modulation/coding for other semantic modalities.Source: 3) Large multimodal model · p. 12

4) World model:

  • A world model combines an environment simulator that predicts future states and rewards with a controller that trains in simulation before acting in the real environment.World models support simulation and training in virtual environments, unlike digital twins’ real-time replication focus.Source: 4) World model · p. 12
  • World models support long-term planning, resource allocation, continuous perception, and embodied tasks such as robot navigation and environmental monitoring.They can simulate strategy outcomes and monitor changing environmental variables for decision-making.Source: 4) World model · p. 12
  • Communication work includes the AGI-native wireless-system vision, which places a world model between perception and action planning and proposes three user- and agent-oriented use cases.The use cases include next-generation digital twins, cognitive avatars, and holographic transmission for a brain-level metaverse experience.Source: 4) World model · p. 13

C. Training of LAMs for communications

1) Pre-training of LAMs for communications:

The section describes pre-training as a way to acquire general features from unlabeled data, followed by fine-tuning that adapts pretrained LAMs to communication tasks. It surveys learning strategies, efficient adaptation methods, and telecom instruction tasks.

1) Pre-training of LAMs for communications:

  • Pre-training on large unlabeled datasets aims to learn universal features, improve communication-task performance and knowledge transfer, and reduce reliance on labeled data.The surveyed approaches include self-supervised learning and multi-task learning.Source: 1) Pre-training of LAMs for… · p. 13
  • Self-supervision creates training signals through transformations or masking, while multi-task learning shares parameters across related tasks to use their relationships.Multi-task models can use shared layers for common features and task-specific layers for individual objectives.Source: 1) Pre-training of LAMs for… · p. 13
  • Distributed training, learning-rate scheduling, and gradient clipping are presented as strategies to improve pre-training efficiency or stability.Examples include data and model parallelism, cosine annealing or cyclic schedules, and absolute-value or norm-based clipping.Source: 1) Pre-training of LAMs for… · p. 13

2) Fine-tuning of LAMs for communications:

  • Fine-tuning adapts a pretrained LAM to communication datasets; telecom instruction tuning pairs natural-language instructions with responses to improve task handling.The instruction dataset is generated using LLMs such as GPT-4 and LLaMA-3 from telecom documents.Source: 2) Fine-tuning of LAMs for… · pp. 13–14
  • Telecom instruction tasks span question answering, technical-document classification, mathematical modeling, code generation, code completion, and code summarization.Modeling tasks include generating communication equations from textual system descriptions.Source: 2) Fine-tuning of LAMs for… · p. 13
  • Parameter-efficient options include LoRA, adapters, BitFit, and prefix tuning, which reduce updates by changing low-rank weights, adding small modules, updating biases, or tuning prefixes.Adapters keep pretrained parameters fixed; prefix tuning also leaves original model weights fixed.Source: 2) Fine-tuning of LAMs for… · p. 14

3) Alignment of LAMs for communications:

The section describes alignment methods for improving the fit between LAM responses and human preferences, then introduces communication evaluation datasets and task categories. It covers both feedback-based training approaches and evaluation scope.

3) Alignment of LAMs for communications:

  • Alignment tuning targets undesirable post-fine-tuning responses, including repetition, overly short replies, and irrelevant content.The section presents alignment as a way to guide responses toward greater accuracy and reasonableness.Source: 3) Alignment of LAMs for communications · p. 14
  • RLHF uses human feedback to train a reward model, then applies reinforcement learning to optimize a LAM’s responses.The workflow includes constructing an environment and agent, collecting feedback, reward modeling, and reinforcement training.Source: 3) Alignment of LAMs for communications · p. 14
  • RLAIF substitutes AI-generated evaluations for human feedback, while PPO constrains policy updates and DPO directly optimizes outputs against preferences without a reward model.The passages describe these as alternatives to RLHF for alignment.Source: 3) Alignment of LAMs for communications · pp. 14–15

D. Evaluation of LAMs for communications

  • TeleQnA contains 10,000 telecommunications questions and answers drawn from standards and research articles, with human input used in its automated question-generation process.The dataset is proposed as a benchmark for assessing LLM telecommunications knowledge.Source: D. Evaluation of LAMs for communications · p. 15
  • The evaluation framework covers communication Q&A, tool learning, communication modeling, and code design.These categories assess knowledge, tool selection and use, mathematical problem-solving, and communication-related coding.Source: D. Evaluation of LAMs for communications · p. 15

1) Communication Q&A:

The section outlines communication evaluations for question answering, tool use, mathematical modeling, and code design, then introduces chain of thought as a structured reasoning optimization. These evaluations assess task-specific correctness alongside reasoning or operational capabilities.

1) Communication Q&A:

  • Communication Q&A evaluates responses to multiple-choice and open-ended questions from communication literature, patents, and books against ground-truth answers.Assessment can include accuracy, comprehension, reasoning, precision, recall, and F1 score.Source: 1) Communication Q&A · p. 15

2) Communication tool learning:

  • Tool-learning evaluation tests whether LAMs select appropriate communication tools and use individual or combined tools to address tasks.The passages cite channel-model code use and multi-tool benchmarks as examples.Source: 2) Communication tool learning · p. 15

3) Communication modeling:

  • Communication modeling tests equation completion and compares predicted mathematical expressions with standard answers for accuracy and equation consistency.The evaluation also examines reasoning about communication principles.Source: 3) Communication modeling · pp. 15–16

4) Communication code design:

  • Code-design evaluation covers generating, completing, and analyzing communication code, including signal-processing and network-protocol tasks.Generated code is assessed for accuracy, completeness, and logical correctness; code analysis also tests communication-programming understanding.Source: 4) Communication code design · p. 16

1) Chain of thought:

  • Chain of thought decomposes complex tasks into sequential reasoning steps, producing a decision or answer that can be checked for correctness and reliability.The paper frames CoT as suitable for multi-step reasoning and comprehensive analysis.Source: 1) Chain of thought · p. 16
  • Communication applications include CoT for FPGA development and FFT implementation [117], distributed-agent collaboration [118], and wireless tasks such as power allocation and spectrum sensing [119].The cited work reports that CoT prompts help decompose tasks; the passages do not establish a single shared evaluation result.Source: 1) Chain of thought · p. 16

2) Retrieval-augmented generation:

The section describes retrieval-augmented generation (RAG) and agentic systems as ways to extend LAM capabilities for communication tasks, then notes unresolved deployment and reliability challenges. It also introduces LAMs as tools for increasingly complex physical-layer design.

2) Retrieval-augmented generation:

  • RAG retrieves relevant documents, combines them with the input question, and uses the fused input to generate more informed answers.Communication examples include systems for 3GPP document question answering, 6G service specifications, and cross-layer tasks.Source: 2) Retrieval-augmented generation · pp. 16–17

3) Agentic system:

  • Agentic systems plan subtasks, execute and adapt actions, then validate outputs and incorporate feedback for later improvement.Communication examples include wireless network management, edge-server collaboration, and generative semantic communication [126].Source: 3) Agentic system · p. 17

1) Summary:

  • The chapter surveys LAM architectures, model classes, training from pretraining through alignment, evaluation, and optimization techniques.It presents this overview as a foundation for LAM applications and future development in communication.Source: 1) Summary · pp. 17–18

2) Lessons learned:

  • Transformers, diffusion, and Mamba show modeling and reasoning capabilities, but face complexity, convergence, training, evaluation, and deployment challenges; CoT, RAG, and agentic systems still have stability, consistency, and efficiency limits.The difficulties are especially noted for resource-constrained environments, multimodal tasks, and real-time communication.Source: 2) Lessons learned · p. 18

III. LAMS FOR PHYSICAL LAYER DESIGN

  • LAMs and generative AI models are presented as tools for addressing increasingly complex physical-layer design challenges.Source: §III · p. 18

A. Channel and beam prediction based on LAMs

The section surveys LAM applications in channel and beam prediction, automated physical-layer design, and resource allocation. It also identifies constraints including weak interpretability, costly data requirements, limited generalization, computational overhead, and insufficient stability.

A. Channel and beam prediction based on LAMs

  • CSI-LLM models variable-length channel histories for multi-step prediction [131], LLM4CP uses channel modules and cross-modal transfer for TDD/FDD prediction, and an LLM beam predictor reports robustness in dynamic settings [132].The beam-prediction comparison is against traditional LSTM models.Source: A. Channel and beam prediction based… · p. 18

B. Automated physical layer design based on LAMs

1) Summary:

  • The chapter characterizes LAMs as promising for channel estimation, blind channel equalization, and physical-layer automation through reasoning and multitask learning.Source: 1) Summary · p. 19

2) Lessons learned:

  • Physical-layer optimization is constrained by limited interpretability and a need for high-quality annotated data that is costly to acquire and label.The passage identifies interpretability and data acquisition as future research directions, particularly for high-reliability communication.Source: 2) Lessons learned · p. 19

A. Computational resource allocation

  • For computational allocation, a diffusion-based ASP-selection method is combined with deep reinforcement learning, while an MoE-and-LLM method uses reasoning for expert selection and decision weighting.The latter is evaluated on maze navigation and network-service-provider utility tasks.Source: A. Computational resource allocation · p. 19

B. Spectrum resource allocation

  • A dialogue-based LLM-and-RAG agent builds system models and applies MoE PPO for spectrum decisions; a GDM-and-TD3 framework with Hungarian RU allocation is simulated for 802.11ax offloading and spectrum allocation [141].The latter reports improvements in bandwidth utilization, latency, and energy consumption in simulations.Source: B. Spectrum resource allocation · p. 19

C. Energy resource optimization

  • Simulations of an AIGC-service framework for energy allocation showed reduced energy consumption and improved content quality and transmission efficiency.The framework dynamically selects service providers using a deep reinforcement learning algorithm [143].Source: C. Energy resource optimization · pp. 19–20

IV. LAMS FOR RESOURCE ALLOCATION AND OPTIMIZATION

  • LAM resource-allocation methods can improve efficiency, but dynamic, resource-constrained settings may yield infeasible strategies, and complex inference can impose overhead that misses real-time needs.Energy scheduling also faces stability and interpretability challenges.Source: 1) Summary · p. 20

2) Lessons learned:

LAMs are surveyed for network design and management and for edge-device training and inference. The section emphasizes feasibility, stability, heterogeneous-data, real-time, compute, and memory constraints.

A. Network design

  • A diffusion model learns design intent to generate customized network solutions, while ChatNet uses domain-adapted LLMs and external tools for design, diagnosis, configuration, and security.A wireless multi-agent GAI network is also proposed for autonomous networking [147].Source: A. Network design · p. 20

B. Network management

  • NetLM uses multimodal representations to analyze packet sequences and network dynamics, while an LLM-and-multi-agent framework maps user intents to slice requirements and supports cross-domain slice management [149].A LAM-enabled architecture also uses edge computing for high-concurrency operations and resource scheduling.Source: B. Network management · p. 20

1) Summary:

  • The chapter summarizes LAM use in network architecture design and slicing management, including processing 6G data for intelligent network operations.Source: 1) Summary · p. 20

2) Lessons learned:

  • Network designs generated under multiple constraints still face feasibility and stability challenges, while management systems struggle with heterogeneous data at scale and real-time requirements.Source: 2) Lessons learned · p. 20

VI. LAMS FOR EDGE INTELLIGENCE

  • Edge-LLM combines layered unified compression, adaptive layer tuning, and hardware scheduling to address edge-device memory and computation overhead.Other approaches include federated fine-tuning under memory-bandwidth limits [155] and edge-terminal collaboration.Source: A. Edge training and application of LAMs · p. 21

A. Edge training and application of LAMs

  • EdgeShard distributes LLMs across devices and uses dynamic programming to select devices and model partitions, reporting 50% lower latency and doubled throughput.The passage also describes edge training, deployment, and inference approaches including compression, caching, and distributed learning.Source: A. Edge training and application of LAMs · p. 21

B. Edge resource scheduling meets LAMs

The section reviews LAM methods for edge resource scheduling, federated learning, and semantic communication. It describes deployment and data constraints alongside systems that adapt models, resource use, and transmitted semantic content.

B. Edge resource scheduling meets LAMs

  • Edge devices’ compute and storage limits motivate joint optimization of offloading, computation, and storage resources for LAM-based tasks.Examples include LAMBO’s embeddings and asymmetric encoder-decoder for offloading decisions and a Stackelberg-game framework for resource allocation.Source: B. Edge resource scheduling meets LAMs · p. 21

C. Federated learning of LAMs

  • Federated learning combines local training and privacy benefits with LAM capabilities for more complex tasks; FwdLLM uses backpropagation-free training with LoRA and adapters on commercial mobile devices [162].Personalized semantic communication and wireless federated tuning are also described [163] [113].Source: C. Federated learning of LAMs · pp. 21–22

1) Summary:

  • The chapter summarizes edge training and deployment, resource scheduling, and federated LAM learning as ways to enhance edge performance and support tasks without centralized data.Source: 1) Summary · p. 22

2) Lessons learned:

  • Edge deployment remains constrained by compute and memory, while federated LAM learning faces resource, data-heterogeneity, and personalization challenges.The section identifies efficient collaboration, robust optimization, and privacy-preserving personalization as research priorities.Source: 2) Lessons learned · p. 22

VII. LAMS FOR SEMANTIC COMMUNICATION

  • The semantic-communication discussion frames LAMs as shifting communication beyond signal transmission toward information comprehension.Source: §VII · p. 22

A. LLM-based semantic communication systems

  • LLM-based systems include physical-layer encoding and decoding without extra retraining [167], talking-face video reconstruction using text and knowledge bases [168], and underwater-image prioritization that reduces data size to 0.8% of the original [169].Source: A. LLM-based semantic communication… · p. 22

B. Other LAM-based semantic communication systems

  • Other LAM semantic systems use cross-modal memory and SNR-aware coding, MLLM-assisted encoding for out-of-distribution images, and LVM-driven 3D semantic extraction and compression [56].Additional proposals include shared-knowledge multiuser communication [176], Mamba-based multimodal communication, and personalized multimodal alignment and recovery.Source: B. Other LAM-based semantic… · p. 23

C. Summary and lessons learned

The section reviews LAM applications in semantic communication, agentic systems, and emerging fields, while identifying open deployment and data challenges. Examples range from personalized healthcare to agent-driven communication and network applications.

1) Summary:

  • LAMs can reduce communication overhead and improve efficiency, semantic expression, and understanding in semantic communication.Direct physical-layer encoding and decoding can perform well, but computational complexity remains a real-time deployment bottleneck in resource-constrained settings, and alignment, ambiguity, and bandwidth optimization remain open issues.Source: 1) Summary · p. 23, 2) Lessons learned · p. 23

VIII. LAM-BASED AGENTIC SYSTEMS

  • Agentic systems are presented as a way to improve semantic transmission efficiency and edge-device resource allocation.The chapter also describes them as using data analysis and processing capabilities to address communication-system challenges and support more efficient information transmission.Source: §VIII · p. 23, 1) Summary · p. 24

A. Agentic systems based on LLMs

  • LLM-agent examples span mobile-edge split learning and model caching, multi-agent communication evaluation, and network-slicing management [126].Other proposals use multi-agent LLMs for network solutions [147], GAI agents with RAG for MIMO design [135] and satellite network design, and automated agents for base-station siting [180].Source: A. Agentic systems based on LLMs · pp. 23–24

B. Agentic systems based on other GAI models

  • Agent systems using other GAI models include reinforcement-learning-based semantic communication for remote monitoring and a proposed 6G agent architecture validated by a prototype.The prototype identifies energy efficiency, security, and agent-customized communication as challenges.Source: B. Agentic systems based on other GAI… · p. 24

2) Lessons learned:

  • Mobile-device compute limits make local and collaborative LAM-agent deployment inefficient, while missing domain knowledge and high-quality communication data constrain multi-agent reasoning in advanced 6G tasks.Caching and task partitioning can improve resource use, but the systems still struggle with high concurrency and low latency.Source: 2) Lessons learned · p. 24

IX. LAMS FOR EMERGING APPLICATIONS

  • Emerging applications reviewed include smart healthcare, carbon-emissions optimization, digital twins, AIoT, integrated satellite-aerial-terrestrial networks, and UAVs.The review presents these as areas where LAMs are being combined with emerging applications.Source: §IX · p. 24

A. Smart healthcare

  • The openCHA LLM framework is presented as a route to personalized smart-healthcare services and medical consultation [185].It integrates external data sources, knowledge bases, and AI analysis models to address personalization, multimodal processing, and real-time knowledge-updating limitations in existing healthcare LLMs.Source: A. Smart healthcare · pp. 24–25

B. Carbon emissions

The section surveys LAM applications in carbon optimization, digital twins, AIoT, and integrated networks, then discusses data and domain-knowledge barriers. It also notes that data quality, privacy, and security concerns can constrain application performance and safety.

B. Carbon emissions

  • A GAI-based low-carbon AIoT solution applies GANs, RAG, and generative diffusion models to resource allocation across energy, data-center, and mobile-edge settings [186].The described applications include renewable-energy use, ICT equipment and cooling management, and power reduction with IRS deployment and semantic communication in mobile-edge networks.Source: B. Carbon emissions · p. 25

C. Digital twins

  • LLM-Twin combines a digital-twin semantic network, small-to-giant model collaboration, and a native security strategy for communication and multimodal processing in digital-twin networks.The passage reports numerical experiments and case studies validating feasibility.Source: C. Digital twins · p. 25

D. Artificial intelligence of things

  • LLMind converts natural-language instructions into finite-state-machine device-control scripts and accumulates experience to support responsive IoT collaboration.It combines LLMs with domain-specific AI modules to automate tasks and coordinate IoT devices.Source: D. Artificial intelligence of things · p. 25

E. Integrated satellite, aerial, and terrestrial networks

1) Summary:

  • Across the surveyed applications, LAMs are described as supporting personalized care, sustainability, industrial automation, IoT collaboration, and networking decisions.The review also identifies security and privacy risks where data tampering or model attacks could affect physical systems, ISATNs, or UAV communications.Source: 1) Summary · p. 25, 2) Lessons learned · p. 26

1) Lack of high-quality communication data:

  • High-quality, diverse communication data are difficult to obtain because collection is costly, privacy-constrained, sparsely labeled, and unable to cover dynamic conditions.Data scarcity can hinder LAM applications; data augmentation, self-supervised learning, and GANs are proposed to expand data and reduce reliance on high-quality labels.Source: 1) Lack of high-quality communication… · p. 26, 2) Lessons learned · p. 26

2) Lack of structured communication knowledge:

  • LAMs can miss structured communication knowledge, including physical constraints such as attenuation, interference, and noise, affecting tasks like interference cancellation and channel modeling.Physics-informed networks and knowledge graphs are proposed for integrating communication principles and structuring domain knowledge.Source: 2) Lack of structured communication… · p. 26

3) Generative hallucination in communication:

The section catalogs limitations affecting LAM reliability, adaptability, interpretability, and deployment in communication systems. The review concludes that addressing these challenges is important to realizing intelligent, efficient, and secure communication services.

3) Generative hallucination in communication:

  • Communication hallucinations include incorrect factual content and failures to follow instructions, potentially degrading signal-quality prediction, network optimization, and user experience.Proposed responses include physical constraints, traditional optimization, ensemble decisions, and specialized detection and mitigation algorithms.Source: 3) Generative hallucination in… · p. 26

4) Limitations of reasoning ability:

6) Adaptability in dynamic environments:

  • Dynamic changes in network topology, channel conditions, and user demands can make LAM adaptability a practical bottleneck.Online, continual, multi-task, and meta-learning are proposed to support adaptation and knowledge transfer.Source: 6) Adaptability in dynamic environments · p. 27

7) Diversity of communication tasks:

  • Communication tasks vary in objectives, constraints, and optimization needs, complicating the design of architectures that can flexibly serve them.Task-specific models, mixture-of-experts, and transfer learning are proposed as approaches for diverse task requirements.Source: 7) Diversity of communication tasks · p. 27

8) Resource constraints at the edge:

  • Limited compute, memory, and energy at mobile, edge, and IoT devices make direct LAM deployment prone to performance degradation and increased latency.Distillation, compression, and hardware acceleration are proposed to improve efficiency in constrained settings.Source: 8) Resource constraints at the edge · p. 27

9) High inference latency:

  • Large model architectures and computational demands can cause high inference latency, challenging real-time communication that requires low latency and high throughput.Operator fusion and speculative sampling are described as ways to reduce inference latency while maintaining accuracy as a key challenge.Source: 9) High inference latency · p. 27

10) Security and privacy:

  • Centralized pretraining, vulnerable data transmissions, and adversarial attacks expose LAM-based communication to privacy, security, and network-stability risks.Federated learning, homomorphic encryption, and secure multi-party computation are proposed to reduce data exposure or protect data in untrusted environments.Source: 10) Security and privacy · pp. 27–28

XI. CONCLUSION

  • The review identifies data, knowledge, hallucination, reasoning, interpretability, adaptability, task-diversity, edge-compute, latency, and security challenges as constraints on practical LAM deployment in communication.It proposes potential solutions and concludes that LAMs are expected to support more intelligent, efficient, and secure services in future networks.Source: §XI · p. 28
Loading 2505.03556v1…