PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities

Hao Zhou, Chengming Hu, Ye Yuan, Yufei Cui, Yili Jin, Can Chen, Haolun Wu, Dun Yuan, Li Jiang, Di Wu, Xue Liu, Charlie Zhang, Xianbin Wang, Jiangchuan Liu

arXiv:2405.10825v2eess.SYcs.LG

TL;DR

Existing telecom LLM studies often address isolated aspects, leaving a need for a broader synthesis across fundamentals, techniques, and applications. This survey organizes the field across generation, classification, optimization, and prediction, covering nearly 20 telecom scenarios as a roadmap for researchers.

  • Problem

    Existing studies focus on specific aspects of telecom LLMs, motivating a broader survey spanning fundamentals, techniques, and application scenarios.

    Source: §I · p. 3

  • Method

    The paper synthesizes LLM fundamentals, deployment approaches, and techniques for telecom generation, classification, optimization, and prediction problems.

    Source: §I · pp. 3–4

  • Results

    The survey covers nearly 20 telecom application scenarios and LLM-inspired techniques, presenting a roadmap for applying LLMs to telecom tasks.

    Source: §I · p. 4

  • Takeaways & Limitations

    The paper provides a broad reference for researchers studying LLM-enabled telecom networks across diverse task types and deployment settings.

    Source: §I · pp. 3–4

Abstract

from arXiv · show

Large language models (LLMs) have received considerable attention recently due to their outstanding comprehension and reasoning capabilities, leading to great progress in many fields. The advancement of LLM techniques also offers promising opportunities to automate many tasks in the telecommunication (telecom) field. After pre-training and fine-tuning, LLMs can perform diverse downstream tasks based on human instructions, paving the way to artificial general intelligence (AGI)-enabled 6G. Given the great potential of LLM technologies, this work aims to provide a comprehensive overview of LLM-enabled telecom networks. In particular, we first present LLM fundamentals, including model architecture, pre-training, fine-tuning, inference and utilization, model evaluation, and telecom deployment. Then, we introduce LLM-enabled key techniques and telecom applications in terms of generation, classification, optimization, and prediction problems. Specifically, the LLM-enabled generation applications include telecom domain knowledge, code, and network configuration generation. After that, the LLM-based classification applications involve network security, text, image, and traffic classification problems. Moreover, multiple LLM-enabled optimization techniques are introduced, such as automated reward function design for reinforcement learning and verbal reinforcement learning. Furthermore, for LLM-aided prediction problems, we discussed time-series prediction models and multi-modality prediction problems for telecom. Finally, we highlight the challenges and identify the future directions of LLM-enabled telecom networks.

I. INTRODUCTION · II. RELATED SURVEYS

The paper surveys LLM-enabled telecom networks from fundamentals through generation, classification, optimization, prediction, applications, challenges, and future directions. It distinguishes its comprehensive, technically detailed coverage from prior surveys by incorporating prompting, deployment, novel LLM-inspired techniques, and broad telecom scenarios.

I. INTRODUCTION

  • 6G’s 107/km2 connection densities, lower than 0.1 ms latency, and integrated functions substantially increase management complexity across configuration, coding, troubleshooting, optimization, and prediction.The architecture includes core, transport, edge, and radio access networks, with satellite communications also integrated for ubiquitous connectivity.Source: §I · p. 1
  • LLMs offer telecom-domain question answering, troubleshooting reports, project code, network configuration, and multimodal processing for sensing and wireless-environment understanding.These capabilities leverage in-context learning, step-by-step learning, and instruction following [12].Source: §I · p. 3
  • Applying general-domain LLMs to telecom can perform poorly because telecom requirements are specialized and scarce, while fine-tuning, dataset curation, and multi-step prompting require dedicated design.Telecom tasks span signal transmission, protocols, architectures, devices, and standards, and coding may involve multiple reasoning steps [13], [14].Source: §I · p. 3

II. RELATED SURVEYS

  • Unlike prior surveys focused on particular aspects such as edge intelligence or grounding, this work comprehensively analyzes LLM fundamentals, deployment strategies, applications, and LLM-enabled optimization and prediction.It explicitly distinguishes text-only LLM inputs and outputs from multimodal LLM applications.Source: §I · p. 3, §II · p. 4

I. INTRODUCTION

  • The survey organizes telecom applications into generation, classification, optimization, and prediction, including code and configuration generation, security and multimodal classification, reward design, verbal reinforcement learning, and time-series prediction.Prediction coverage includes frozen and fine-tuned models, foundation-model pre-training, and multimodal LLMs.Source: §I · pp. 3–4
  • The survey covers LLM fundamentals, generation, classification, optimization, prediction, nearly 20 telecom application scenarios, and future research directions as a roadmap for LLM-enabled telecom networks.It spans model architecture, pre-training, fine-tuning, inference, evaluation, deployment, and telecom applications.Source: §I · pp. 3–4

II. RELATED SURVEYS

  • Prior surveys cover LLM training, deployment, satellite-aerial-terrestrial, secure, semantic, vehicular, and sensing applications, but this work systematically extends coverage across question answering, troubleshooting, coding, configuration, security, and classification.Table I compares the surveyed topics, while Table II summarizes general and domain-specific LLMs across fields including healthcare, finance [10], time series [47], autonomous driving [48], and recommendation [49].Source: §II · pp. 4–5, Table I · p. 5
  • Compared with existing surveys, the paper adds chain-of-thought and step-by-step planning, systematic deployment analysis, automated reward-function design, time-series one-model-for-all prediction, and detailed application-level frameworks and prompts.Existing work often emphasizes general system designs, whereas this survey provides technical details on pre-training steps and prompt designs; automated reward design can achieve performance comparable to manual designs [42]–[44], and time-series LLMs enable one-model-for-all prediction.Source: §II · pp. 4–5

III. LLM FUNDAMENTALS … 3) Decoder-only architecture:

Section III surveys LLM fundamentals, covering transformer-based architectures, training and use procedures, evaluation, telecom deployment, and domain-specific application perspectives. It distinguishes encoder-only, encoder-decoder, and decoder-only designs, with causal decoders dominating contemporary LLMs.

III. LLM FUNDAMENTALS

  • LLM fundamentals encompass architecture, pre-training, fine-tuning, inference and utilization, evaluation, telecom deployment, and telecom-domain training or fine-tuning.Deployment settings include central clouds, network edges, and mobile devices.Source: §III · p. 5

A. Model Architecture

  • Transformers tokenize inputs and apply embeddings and positional encodings, while attention captures global input-output dependencies.Vanilla transformers comprise encoder and decoder components, and attention variants include multi-head, multi-query [51], and grouped-query attention.Source: A. Model Architecture · pp. 5–6

1) Encoder-only architecture:

  • Encoder-only models extract language features for understanding tasks such as classification, exemplified by BERT [53] and its RoBERTa [54] and ALBERT [55] variants.BERT uses masked-language-model and next-sentence-prediction objectives; ALBERT reduces parameters to accelerate training.Source: 1) Encoder-only architecture · p. 6

2) Encoder-decoder architecture:

  • Encoder-decoder architectures pass encoder self-attention keys and values to the decoder for cross-attention, supporting sequence-to-sequence transfer and denoising pretraining.T5 reformulates text tasks into text-to-text formats, while BART [57] reconstructs noise-corrupted text by combining bidirectional encoding with causal decoding.Source: 2) Encoder-decoder architecture · p. 6

3) Decoder-only architecture:

  • Decoder-only architectures use unidirectional attention, processing prefix and output tokens identically so each output attends only to itself and preceding tokens.This design specializes in autoregressive generation from prior context.Source: 3) Decoder-only architecture · p. 6
  • Decoder-only models divide into causal and non-causal decoders: causal variants restrict every token to past tokens and itself, whereas non-causal variants bidirectionally process prefixes before sequential generation.Causal decoders dominate popular LLMs including GPT, PaLM, and LLaMA [60], while non-causal decoders [61] resemble encoder-decoder frameworks.Source: 3) Decoder-only architecture · p. 6
  • Telecom applications can pre-train an LLM from scratch on telecom data, fine-tune a general-domain model for specific tasks, or prompt a general-domain model directly.Sections III-B, III-C, and III-D introduce pre-training, fine-tuning, and prompting procedures, respectively.Source: 3) Decoder-only architecture · p. 6

B. LLM Pre-training … 2) Chain-of-thought (CoT) prompting:

The paper surveys LLM development from pre-training and fine-tuning to prompt-based inference, emphasizing the data, scaling, alignment, and reasoning techniques that enable telecom applications. It covers ICL and CoT prompting as training-free methods for adapting LLMs to new tasks and complex reasoning.

B. LLM Pre-training

  • Pre-training predicts the next word from large datasets, with general and specialized data supporting broad language understanding, reasoning, and domain-specific knowledge.Collected data require quality filtering, de-duplication, privacy redaction, and tokenization [66] before training; specialized sources include scientific literature and programming repositories.Source: B. LLM Pre-training · p. 6, 1) Dataset collection and preprocessing · pp. 6–7

2) Model training:

  • Large-scale pre-training relies on substantial batches, warm-up followed by cosine-decay learning rates, and distributed methods including 3D parallelism and ZeRO.Data, pipeline, and tensor parallelism distribute data, layers, and tensors across GPUs, while ZeRO reduces memory redundancy by retaining only portions of model data on each GPU.Source: 2) Model training · p. 7

C. LLM Fine-tuning

  • Fine-tuning adapts pre-trained LLM parameters to domain-specific tasks, with instruction tuning using natural-language instances to improve generalization to unseen and multilingual tasks [71].Instruction instances combine task descriptions, optional inputs, outputs, and demonstrations; quality and diversity matter, while scaling instructions improves generalization only up to a point [74].Source: C. LLM Fine-tuning · p. 7, 1) Instruction tuning · p. 7

2) Alignment tuning:

  • Alignment tuning uses human feedback to reduce harmful, biased, or misleading outputs, with RLHF combining supervised fine-tuning, reward-model training, and reinforcement learning.DPO removes the reward-model stage [78], while Constitutional AI [79] and RLAIF [80] generate preference data with AI feedback to address costly and limited human preference data.Source: 2) Alignment tuning · pp. 7–8

D. LLM Inference and Utilization by Prompting

  • Prompt engineering generates desired outputs without extra training, offering an immediate and training-free approach for telecom applications.The surveyed techniques include ICL, CoT prompting, complex planning, and self-refinement with iterative feedback.Source: D. LLM Inference and Utilization by… · p. 8

1) In-context learning (ICL):

  • ICL uses task descriptions and demonstrations in structured prompts, enabling LLMs to recognize tasks and learn new solving strategies from context.Demonstration selection, formatting, and ordering are critical, and larger models show stronger ability to learn beyond prior knowledge from examples [83], [84],, [87].Source: 1) In-context learning (ICL) · p. 8

2) Chain-of-thought (CoT) prompting:

  • CoT prompting extends ICL with intermediate reasoning steps, improving LLM performance on arithmetic, commonsense, and symbolic reasoning tasks.Reasoning steps can be manually annotated or elicited with phrases such as “Let’s think step by step,” making implementation easier.Source: 2) Chain-of-thought (CoT) prompting · p. 8
  • CoT can suffer from incorrect reasoning and instability, motivating diverse reasoning paths, advanced generation, verification, and tree- or graph-structured reasoning.Its benefits are strongest for LLMs above 10B parameters and tasks requiring detailed solutions, whereas it may underperform on simpler tasks or when traditional prompting already works.Source: 2) Chain-of-thought (CoT) prompting · p. 8

3) Planning for complex task solving: … 3) On-device deployment::

The section presents planning and self-refinement methods for complex LLM task solving, then defines evaluation dimensions and deployment strategies for telecom networks. It contrasts cloud, edge, and on-device deployment according to resource demands, latency, bandwidth, and collaboration requirements.

3) Planning for complex task solving:

  • Prompt-based planning decomposes complex tasks into sub-tasks, executes action sequences with environmental feedback, and refines plans through reasoning, backtracking, or memorization.The framework comprises a task planner, plan executor, and environment; plans may be natural-language sequences or executable programs.Source: 3) Planning for complex task solving · p. 8

4) Self-refinement with iterative feedback:

  • Self-refinement uses actionable feedback and iterative revision to improve initial LLM outputs, including TrafficLLM’s prediction, feedback-generation, and prediction-refinement cycle.The same LLM generates outputs, evaluates them, and refines performance until convergence or through refinement demonstration prompts.Source: 4) Self-refinement with iterative… · pp. 8–9, Figure 3 · p. 9

E. Evaluation metrics of LLM

  • LLM evaluation should jointly consider accuracy, hallucination, efficiency, and human alignment to assess reliability, resource demands, factual consistency, and conformity with human values.ROUGE and BERTScore [94] can be suboptimal for factual consistency, while AlignScore [96] targets hallucination detection and manual evaluation supports alignment assessment.Source: E. Evaluation metrics of LLM · p. 9

F. LLM Deployment in Telecom Networks

  • Telecom deployment must address LLM computational, storage, and latency burdens across cloud, edge, on-device, cache-based, and cooperative strategies.GPT4 has 1.76 trillion parameters and a 45 GB model size, while fine-tuning a 7-billion-parameter model can take nearly 3 hours on an 8×80GB A100 machine [110].Source: F. LLM Deployment in Telecom Networks · pp. 9–10, Table III · p. 10

1) Cloud deployment:

  • Cloud deployment supplies substantial computation and integrates easily with telecom architectures, but data transmission causes long response times and high bandwidth costs, limiting latency-critical applications.Cloud-hosted LLMs can support edge AI by returning task plans and model configurations to network-edge devices.Source: 1) Cloud deployment · p. 10

2) Network edge deployment::

  • Network-edge deployment reduces response time and bandwidth use, while parameter sharing, model compression, parameter-efficient finetuning, split edge learning, and quantized training address limited resources.Edge cloud and base-station deployments remain more resource-constrained than central-cloud deployment.Source: 2) Network edge deployment · p. 10

3) On-device deployment::

  • On-device deployment enables fast, customized responses but remains resource-intensive and immature; split end-edge-cloud systems pair sub-10B mobile models with larger edge models for complex tasks.LLAMA-7B is assumed operable on mobile devices, while GPT-4 is placed at the edge; Apple reports 20 times faster flash-memory inference than GPU use under limited DRAM [106].Source: 3) On-device deployment · p. 11, 4) Cache-based deployment · p. 11

4) Cache-based deployment: … B. Domain Knowledge Generation

The surveyed telecom LLM methods span cache-based and cooperative deployment, telecom-domain fundamentals, and generation applications from knowledge access to configurable outputs. Together, these approaches target efficient, reliable telecom deployment while exploiting LLM reasoning, customization, and generation capabilities.

4) Cache-based deployment:

  • Cache-based deployment reduces latency from full model migration but requires complex coordination for model-update synchronization frequency and quantization-bit selection.Source: 4) Cache-based deployment · p. 11
  • 3.2x lower end-to-end latency and 34.3% accuracy improvement than the baseline were achieved by EdgeFM through selective cloud querying and periodic edge-model customization.Edge devices assess uncertainty and network dynamics, upload unseen classes, and transition from frequent cloud queries to customized local models as conditions stabilize.Source: 5) Cooperative deployment · p. 11

5) Cooperative deployment:

  • Cooperative deployment combines small edge models with larger cloud models to retain accuracy while reducing computational-resource dependence, including cloud-edge offload, splitting, and synergy architectures.Service-delay evaluation and task offloading remain open research problems.Source: 5) Cooperative deployment · pp. 11–12

G. Analyses of LLM Fundamentals in the Telecom Domain

  • Prompting methods support real-time and multi-step telecom tasks through in-context learning, chain-of-thought reasoning, and structured planning with feedback and plan refinement.Their effectiveness depends on demonstration design, while resource allocation and multi-agent optimization require carefully designed multi-step reasoning.Source: G. Analyses of LLM Fundamentals in the… · p. 12
  • Telecom LLM evaluation should measure efficiency, accuracy, reliability, hallucination, and ethical responsibility, particularly for mission-critical applications such as URLLC.Long inference times or misleading network-management decisions can make models unsuitable for telecom environments.Source: G. Analyses of LLM Fundamentals in the… · p. 12
  • Telecom deployment must address stringent delay and reliability requirements alongside limited device computation and storage, motivating efficient training, fine-tuning, inference, and storage techniques [16].Fine-tuning general-domain LLMs is more practical than training from scratch, but telecom datasets face complex preprocessing, computational burdens, privacy constraints, and unresolved alignment challenges.Source: G. Analyses of LLM Fundamentals in the… · p. 12

IV. LLM FOR GENERATION PROBLEMS IN WIRELESS NETWORKS

  • LLM generation in wireless networks targets telecom knowledge, code, and network-configuration tasks, motivated by the technology’s broad knowledge and strong generation capability.The survey organizes generation-related studies and opportunities across these application categories in Table IV.Source: §IV · p. 12, Table IV · p. 13

A. Motivations of Using LLM-based Generation for Telecom

1) Understanding telecom domain knowledge: … 1) LLM for code refactoring:

The section surveys how telecom-specific LLM training supports domain-knowledge generation and how LLMs improve telecom code development. It covers datasets, training strategies, question answering, troubleshooting, code generation, refactoring, and validation.

1) Understanding telecom domain knowledge:

  • Telecom knowledge spans signal transmission, network architectures, communication protocols, and standards such as 3G, 4G, 5G, and IEEE 802.11, making domain understanding central to interoperable and secure systems.Domain-specific LLM training can automate knowledge generation and make complex telecom content more accessible to experts and general users.Source: 1) Understanding telecom domain… · p. 14

2) Training LLMs with telecom-specific data:

  • Telecom-specific LLM training requires collecting professional datasets, preprocessing them, and choosing between training from scratch and fine-tuning a general-domain model.Collected sources include technical documents, research papers, standards, TeleQuAD, 185,000 trouble reports, and a large telecom-knowledge dataset [125]; training from scratch can outperform fine-tuning, while fine-tuning is more efficient.Source: 2) Training LLMs with telecom-specific… · pp. 14–15

3) Using LLM to telecom knowledge-related generation tasks:

  • Telecom-domain question answering broadens access to complex standards and technologies, but differing answers from GPT-4 and LLaMA show that domain-specific training is important to avoid misleading users.For the 5G spectrum-layer question, GPT-4 and LLaMA provide different frequency bands, potentially because of different pre-training data sources [117].Source: 3) Using LLM to telecom… · pp. 15–16
  • LLM-based troubleshooting generates and ranks top-K solutions from trouble-report observations, headings, and fault areas, improving response and repair efficiency but requiring verification because the best correct rate in is around 60%.The BERT-based approach in achieved a nearly 55% correct rate, while transfer learning and non-task-specific telecom data improved generalization to unseen reports in.Source: Figure 5 · p. 15, 3) Using LLM to telecom… · p. 16

C. Code Generation

  • LLMs support telecom code generation and repair across Python and Verilog, including an 86.71% repair rate for 286 student programs and reduced Verilog coding time with improved undergraduate and postgraduate code quality [120], [14].These capabilities can reduce human effort in coding, validation, and debugging for telecom scheduling and management projects.Source: C. Code Generation · p. 16

1) LLM for code refactoring:

  • LLM-assisted refactoring targets readability, efficiency, and reliability in evolving wireless systems whose large codebases combine diverse modules and contributors.In OpenWiFi, ChatGPT proposed meaningful names, added comments and explicit wire types, and introduced asynchronous reset handling for faster error responses [14].Source: 1) LLM for code refactoring · p. 16, Figure 6 · p. 17
  • LLM-assisted validation can generate an error-free OpenWiFi testbench, while fine-tuning on GitHub and textbook Verilog data improved coding correct rate by 26% for a specific language [14],.The OpenWiFi validation example did not investigate fine-tuning, which may be prerequisite for effective hardware-description-language generation.Source: 1) LLM for code refactoring · p. 16

2) LLM-aided code generation with multi-step scheduling: · D. Network Configuration Generation · E. Discussions and Analyses

LLM-aided telecom generation benefits from decomposed, multi-step prompting, examples, libraries, and verification. Applications span code and network configuration generation, while correctness, security, interoperability, and validation remain central challenges.

2) LLM-aided code generation with multi-step scheduling:

  • LLM-plus-library prompting achieved 88% coding accuracy for traffic analysis and 78% for network lifecycle management using GPT-4 and NetworkX.The method separates application prompts from code-generation prompts and uses plugins and libraries for task-specific generation.Source: 2) LLM-aided code generation with… · p. 17, E. Discussions and Analyses · p. 20
  • Step-by-step prompting, examples, pseudocode, debugging, and explicit data types improve LLM code generation and can lower professional-knowledge requirements [14],,.Studies report that students can achieve comparable coding quality or reproduce existing network-study results with LLM assistance.Source: 2) LLM-aided code generation with… · p. 18

D. Network Configuration Generation

  • Network configuration automation targets costly, intricate manual operations across switches, routers, servers, and interfaces, but must handle contextual understanding, security, and vendor interoperability.NETBUDDY translates natural-language policies into formal specifications, forwarding information, and vendor-agnostic device configurations, while CloudEval-YAML benchmarks cloud-native YAML generation [24], [123].Source: D. Network Configuration Generation · pp. 18–19
  • VPP achieved 10× leverage translating Cisco configurations to Juniper through 20 automated prompts and 6× leverage implementing no-transit policies across six routers with 12 prompts.VPP combines GPT-4 with Batfish to identify and fix syntax, structural, attribute, topology, and semantic-policy errors.Source: D. Network Configuration Generation · p. 19
  • Network configuration generation frameworks range from direct LLM output to hierarchical multi-LLM pipelines and automated verifier-driven refinement.Verification checks syntax, compilability, and correctness; verifiers can be inserted after each LLM iteration, although human inspection remains necessary.Source: D. Network Configuration Generation · p. 19, Figure 7 · p. 19

E. Discussions and Analyses

  • Step-by-step planning makes complex telecom generation feasible, whereas one-shot prompting is impractical for tasks such as 64-point FFT generation [14].The approach decomposes objectives into stages and uses progressively generated components, examples, and suggestions.Source: 2) LLM-aided code generation with… · p. 18, E. Discussions and Analyses · p. 19
  • LLM generation can reduce manual coding, validation, debugging, and literature-review effort, with introductory Python repair improving from 86.71% to 96.50% using few-shot examples [120].LLMs can also abstract network knowledge from textbooks, journals, and specification standards.Source: E. Discussions and Analyses · p. 20
  • LLM-generated telecom code and configurations require automatic validation because outputs may be incorrect, contain wrong numbers or units, or introduce security vulnerabilities.Error-message feedback can support automatic debugging, while configuration validation addresses syntax, semantics, security, and correctness.Source: D. Network Configuration Generation · p. 19, E. Discussions and Analyses · p. 20

V. LLM-ENABLED CLASSIFICATION PROBLEMS … 1) Fine-tuning pre-trained LLMs:

LLMs address telecom classification by processing heterogeneous multimodal data, supporting security, text, image, and encrypted-traffic tasks. In attack detection, domain adaptation enables stronger understanding of cybersecurity language than general-purpose models.

V. LLM-ENABLED CLASSIFICATION PROBLEMS

  • Telecom classification spans attack detection, text, image, and encrypted-traffic problems, where accurate and robust decisions improve network service quality and performance.The section frames LLMs as tools for a broad range of telecom classification applications.Source: §V · p. 21

A. Motivations and Classification Capabilities of LLMs

  • LLMs suit telecom classification because they can process heterogeneous multimodal data, including numerical traffic data, textual security logs, and environmental images.Their capabilities across natural language and computer vision motivate their use as more adaptable alternatives to conventional statistical techniques.Source: A. Motivations and Classification… · p. 21

B. LLM for Telecom Security and Attack Detection

A. Motivations and Classification Capabilities of LLMs

  • LLMs can classify customer feedback and extract useful QoE information from telecom texts, helping operators understand user satisfaction and improve service quality.Relevant text includes customer feedback, standards, technical reports, and publications.Source: A. Motivations and Classification… · p. 21
  • Vision-capable LLMs can combine visual and network data for telecom management, while zero-shot classification enables rapid target-task detection without additional training,,, -.Applications include locating users with base-station cameras, image-to-text generation, object detection, network attack detection, and image classification.Source: A. Motivations and Classification… · p. 21

B. LLM for Telecom Security and Attack Detection

  • Telecom attack detection is difficult because evolving infrastructures expose vulnerabilities while redundant, correlated multimodal measurements and device texts can obscure critical attack patterns.Examples include DoS attacks that reduce availability and MITM attacks that compromise communication integrity.Source: B. LLM for Telecom Security and Attack… · p. 21

1) Fine-tuning pre-trained LLMs:

  • SecureBERT adapts a RoBERTa-style twelve-layer architecture using a 1.1-billion-word cybersecurity corpus and a 50,265-token security-customized vocabulary.The vocabulary contains 17,673 cybersecurity-specific tokens, while 32,592 tokens are shared with RoBERTa.Source: 1) Fine-tuning pre-trained LLMs · p. 23
  • SecureBERT correctly predicted reconnaissance, hijacking, and DDoS in cybersecurity masked-word tasks, whereas RoBERTa produced incorrect predictions.The evaluation used 17,341 sentences extracted from cybersecurity reports.Source: 1) Fine-tuning pre-trained LLMs · p. 23

2) Building security-specific LLMs from scratch: … D. Image Classification

The section surveys LLM applications for telecom security and classification, spanning security-specific models, customer feedback, standards, and wireless-environment images. It highlights strong attack-detection and text-classification results alongside edge-based image processing for beamforming and hand-off decisions.

2) Building security-specific LLMs from scratch:

  • SecurityBERT converts hashed numerical and categorical network features into fixed-length text, fine-tunes with Softmax for attack detection, and uses FalconLLM to assess severity and formulate mitigation procedures.Its pipeline includes data preparation, tokenization, pre-training, fine-tuning, and deployment on EdgeIIoTset, covering five threat categories.Source: 2) Building security-specific LLMs… · pp. 23–24
  • SecurityBERT identifies normal events and 14 attacks with average accuracy 0.98, recall 0.84, and F1-score 0.84, surpassing traditional ML and deep learning models.The attacks include DDoS UDP, DDoS ICMP, SQL Injection, and Vulnerability Scanner.Source: 2) Building security-specific LLMs… · p. 24
  • Security-customized telecom corpora, datasets, and tokenizers can improve LLM performance for cyber threat intelligence, vulnerability analysis, and threat action extraction,.The proposed development process begins by creating relevant training and testing datasets, followed by model fine-tuning.Source: 2) Building security-specific LLMs… · p. 24

C. Text Classification

  • Telecom text-classification applications include user enquiries, trouble reports, standard specifications, and customer feedback, supporting more intelligent and reliable networks.LLMs are applied to text-related classification tasks across diverse telecom documents and user-generated content.Source: C. Text Classification · p. 24, 2) LLM-aided telecom standard… · p. 25

1) Using LLMs for telecom user feedback classification and analyses:

  • Customer-feedback analysis found that 78.3% of complaints concerned weak signal coverage and that 92% of those regions had coverage problems, while AraBERT outperformed conventional CNN and RNN algorithms [142],.LLMs can classify sentiment and topics, analyze customer experience and intent, and support user-intent taxonomy generation, validation, and application [160].Source: 1) Using LLMs for telecom user… · p. 24

2) LLM-aided telecom standard classification:

  • LLM-based 3GPP specification classification achieved 80% accuracy using only 20% of text segments, with longer technical segments significantly improving accuracy [13].The method automatically assigns specifications to existing 3GPP working groups after preprocessing and fine-tuning.Source: 2) LLM-aided telecom standard… · pp. 24–25

D. Image Classification

  • LLMs improve zero-shot image classification through crafted prompts, hierarchical strategies, adjusted temperatures, and transferred knowledge, while also supporting image description and accident-report generation,, [17].These capabilities address complex wireless environments containing pedestrians, vehicles, blockages, and other obstacles.Source: D. Image Classification · p. 25
  • In an edge-assisted wireless-vision framework, base-station cameras send environmental images to an edge LLM for classification and object localization, enabling beamforming and hand-off adjustments.The framework detects vehicles, users, and blockage buildings after image processing at the edge cloud.Source: D. Image Classification · p. 25, Figure 10 · p. 25

E. Encrypted Traffic Classification … A. Motivations and Optimization Capabilities of LLM

LLMs improve encrypted traffic classification and support versatile telecom classification, while offering instruction-following, few-shot, and world-knowledge capabilities for network optimization. These capabilities extend from security monitoring and zero-shot multimodal classification to optimization methods addressing telecom-specific formulation and training challenges.

E. Encrypted Traffic Classification

  • Encrypted traffic classification remains challenging because advanced encryption obscures robust patterns, while conventional and labelled-data-dependent ML methods have limited generalization.Conventional methods rely on certificate-based fingerprint matching, whereas ML-based approaches depend heavily on labelled-data amount and distribution.Source: E. Encrypted Traffic Classification · p. 25
  • BERT-BiLSTM achieves 99.70% accuracy, 99.34% precision, 99.51% recall, and 99.43% F1-score on the 17-category ISCX VPN dataset, surpassing traditional ML methods.BERT captures long-distance traffic context, while BiLSTM integrates forward and backward packet-level semantic features.Source: E. Encrypted Traffic Classification · p. 26
  • 5.4%, 0.2%, and 5.2% improvements over existing state-of-the-art methods are reported by ET-BERT across general encrypted application, encrypted malware, and VPN traffic classification tasks.ET-BERT learns generic traffic representations from large-scale unlabelled encrypted traffic using Datagram2Token preprocessing and pre-training.Source: E. Encrypted Traffic Classification · p. 26
  • ET-BERT’s clean-pretraining assumption creates vulnerability to poisoned backdoors, motivating investigation of toxic tokens constructed within encrypted traffic.Attackers can insert low-frequency words as toxic embeddings that deceive fine-tuned classifiers on specific tasks.Source: E. Encrypted Traffic Classification · p. 26

F. Discussions and Analyses

  • LLMs support telecom security by processing correlated multimodal and heterogeneous data for monitoring and attack detection, while recommending incident response and recovery strategies.This suggests an end-to-end security system spanning status monitoring, detection, response, and recovery.Source: F. Discussions and Analyses · p. 26
  • LLMs provide zero-shot classification for diverse telecom tasks, including base-station camera images, customer comments, object detection, and user localization.They also handle natural-language comments and system-language network logs with less human intervention than manual classification.Source: F. Discussions and Analyses · pp. 26–27
  • LLMs extend vision-aided telecom through image description, image-text transformation, object detection, and classification, with applications in blockage prediction, beamforming and hand-off, and user localization.Pre-trained real-world knowledge supports these tasks without extra training and improves generalization over conventional algorithms.Source: F. Discussions and Analyses · p. 27

VI. LLM-ENABLED OPTIMIZATION TECHNIQUES FOR TELECOM

  • LLM-enabled optimization addresses the theory-application gap across reinforcement learning, convex, heuristic, and black-box methods by simplifying formulation, reducing training demands, and leveraging world knowledge.Telecom optimization is hindered by expert-dependent reward design, trial-and-error metric integration, and standard-form constraints for convex solvers such as CPLEX and LINDO.Source: §VI · p. 27, A. Motivations and Optimization… · pp. 27–28

B. LLM-aided Reinforcement Learning for Network Optimization · 1) Using LLM for reward function design:

The section presents LLM-aided reinforcement learning for telecom network optimization, focusing on automatic reward-function design to reduce trial-and-error and expert burden. It formulates prompt- and feedback-driven reward generation, iterative evaluation, and self-improvement, while noting that telecom deployment remains challenging for complex multi-element scenarios.

B. LLM-aided Reinforcement Learning for Network Optimization

1) Using LLM for reward function design:

  • The LLM-aided MDP extends <S, A, R, T> with prompt inputs L and a mapping function M, defining the reward as F := G(L, M).Prompts can describe the task, objective, states and actions, and example trajectories; M maps LLM output to feedback such as “good” or “bad”.Source: 1) Using LLM for reward function design · pp. 28–29
  • The reward-design workflow inputs task information, generates an initial reward function, trains the reinforcement-learning agent, evaluates outcomes, and iteratively incorporates feedback.The process repeats redesign and training until the desired performance is achieved or the maximum iteration number is reached.Source: 1) Using LLM for reward function design · pp. 29–30
  • In the resource-slicing example, LLM-generated rewards are evaluated against user-defined criteria for URLLC-like and eMBB-like groups, and “bad” results trigger redesign.A 5% drop rate for group 1 exceeded the predefined 1% threshold, producing a “bad” evaluation and detailed feedback for the next iteration.Source: 1) Using LLM for reward function design · p. 30
  • Iterations increased the coefficient for group 1 drop rate from 1 to 10 and adjusted the group 2 throughput coefficient to balance the two groups’ latency and throughput metrics.The revised reward was intended to prevent group 1 user drops while balancing performance across the user groups.Source: 1) Using LLM for reward function design · p. 30
  • Automatic reward design can lower the difficulty of applying reinforcement learning to telecom, but complicated transformations, diverse metrics, and multiple network elements remain unresolved challenges.The telecom application is still an open issue despite demonstrations in robotics and logic games.Source: 1) Using LLM for reward function design · p. 30

2) Verbal reinforcement learning via LLM: · C. LLM as a Black-box Optimizer

The section describes LLM-based verbal reinforcement learning and black-box optimization for telecom networks, emphasizing iterative improvement, reduced modeling effort, and remaining reliability and domain-knowledge challenges. Verbal reinforcement learning uses language feedback and memory, while black-box optimization searches solutions from objective scores without gradients or Hessians.

2) Verbal reinforcement learning via LLM:

  • Verbal reinforcement learning, proposed in, uses LLM agents to explore policies and achieves satisfactory performance across sequential decision-making, coding, and language-reasoning tasks.A radio access network optimization example illustrates its telecom application.Source: 2) Verbal reinforcement learning via LLM · p. 30
  • The verbal reinforcement learning agent combines an LLM actor, evaluator, self-reflection, and short- and long-term memories to generate, assess, refine, and retain network decisions.The actor can use CoT and ReAct [186], while self-reflection converts environmental and evaluator feedback into detailed instructions.Source: 2) Verbal reinforcement learning via LLM · pp. 30–31
  • Compared with conventional reinforcement learning, verbal learning avoids hyperparameter tuning, accepts operator language guidance, and provides reasoning and interpretable explanations for policy performance.Language instructions can guide exploration more efficiently than ε-greedy strategies and do not require operators to have machine-learning knowledge.Source: 2) Verbal reinforcement learning via LLM · pp. 31–32
  • Verbal reinforcement learning remains at an early stage in telecom and may require wireless-domain datasets such as TeleQnA [125] for fine-tuning.The passage notes that domain knowledge can help LLMs understand user demand.Source: 2) Verbal reinforcement learning via LLM · p. 32

C. LLM as a Black-box Optimizer

  • An LLM black-box optimizer searches an unknown objective function without gradients or Hessians, avoiding dedicated optimization-model construction as telecom networks become more complex [177].Existing studies show LLMs can fit unknown loss functions.Source: C. LLM as a Black-box Optimizer · p. 32
  • Black-box optimization iteratively evaluates LLM-generated solutions, stores solution-score pairs, and uses similarities among high-scoring examples for few-shot improvement until satisfaction or the iteration limit.The BS power-control example maximizes average data rate by adjusting two base-station transmission powers.Source: C. LLM as a Black-box Optimizer · p. 32
  • Black-box optimization may produce unreliable results because example quality affects outputs and performance is not guaranteed, conflicting with telecom service-level requirements for stringent solution quality.These quality requirements can obstruct LLM adoption in telecom management.Source: C. LLM as a Black-box Optimizer · pp. 32–33

D. LLM-aided Convex Optimization for Telecom · E. LLM-based Heuristic Algorithm Design · F. Discussions and Analyses

The section presents LLMs as tools for automating convex optimization and designing heuristic algorithms in telecom, while emphasizing that task descriptions, prompts, feedback, and planning remain decisive challenges. These methods can reduce human effort and support near-optimal or optimized network solutions, but complex large-scale problems still require careful decomposition and domain expertise.

D. LLM-aided Convex Optimization for Telecom

  • LLMs can diagnose infeasible constraints and support convex-optimization modelling, code generation, and solving, potentially making expert-dependent telecom optimization more accessible [178] [179].Convex methods offer stable and efficient solutions, but telecom problems often require modelling, transformation, and relaxation because they are non-convex.Source: D. LLM-aided Convex Optimization for… · p. 33
  • A standardized description of problem type, variables, objectives, and solvers lowers LLM comprehension difficulty, while telecom fine-tuning and formulation templates can improve modelling accuracy.LLMs can also exploit Gurobi and cvxpy features, such as gurobi.abs, and use code-fix templates to rerun failed implementations.Source: D. LLM-aided Convex Optimization for… · p. 33
  • GPT-4 solved 41 linear and 11 mixed-integer linear programs with a nearly 0.8 success rate on small-scale problems, improved further by supervised tests and data augmentation [179].The end-to-end system models problems, generates code, invokes solvers, and automatically debugs failures.Source: D. LLM-aided Convex Optimization for… · p. 33, Figure 15 · p. 34

E. LLM-based Heuristic Algorithm Design

  • LLMs can decompose heuristic-algorithm design into requirements, candidate selection, component analysis, stepwise design, pseudocode generation, and motivation explanation.Chain-of-thought decomposition lowers prompt difficulty and improves output performance for swarm-based meta-heuristic design.Source: E. LLM-based Heuristic Algorithm Design · p. 34
  • LLM-aided heuristic design could generate fast-converging, low-complexity algorithms for dynamic telecom control and new network protocols or management policies [30].Heuristic methods are useful for integer-control problems such as RIS phase-shift optimization, but selecting among many algorithms has become harder.Source: E. LLM-based Heuristic Algorithm Design · p. 34

F. Discussions and Analyses

  • Optimization performance depends on precise task descriptions and prompt engineering because LLMs may not understand specialized telecom concepts such as RIS phase-shift-control difficulty.Relevant techniques include CoT, ReAct [186], and zero-shot instruction, while user prompts fully determine heuristic-algorithm outputs.Source: F. Discussions and Analyses · pp. 34–35
  • Feedback mechanisms iteratively improve reward functions and actions, but larger telecom action spaces intensify exploration-exploitation conflicts and require dedicated evaluator and feedback prompts.Using LLM self-improvement while maintaining this balance remains important for optimization performance.Source: F. Discussions and Analyses · p. 35
  • Multi-variable, multi-layer wireless optimization may require decomposition, multiple objectives, and alternating optimization, yet current LLMs still lack sufficient planning capabilities for such tasks.Joint transmission-rate and signal-coverage optimization exemplifies the complexity that remains challenging.Source: D. LLM-aided Convex Optimization for… · pp. 33–34, F. Discussions and Analyses · p. 35

VII. TIME SERIES LLM FOR PREDICTION PROBLEMS … 2) Tokenization mechanisms:

The section motivates time-series LLMs for telecom prediction, formulates zero-shot forecasting, and reviews foundation-model pre-training, tokenization, and backbone architectures. It emphasizes generalization across dynamic telecom data, multimodal context, reduced token costs, and rapid forecasting for 6G networks.

A. Motivations

  • LLMs address nonlinear, non-stationary telecom data by adapting to changing patterns, supporting generalizable prediction across traffic, demand, failures, and service usage.Their in-context learning can map demonstrations to predictions without additional training, while multimodal models can combine sensing data with numerical time series for context-aware forecasting.Source: A. Motivations · pp. 35–37

B. Pre-training Foundation Models for Zero-shot Prediction

2) Tokenization mechanisms:

  • Patching segments each time series into overlapping or separate length-P patches with stride S, reducing input tokens from L to roughly L/S.The mechanism appends S repetitions of the final value before patching and is intended to reduce memory consumption and computational intensity.Source: 2) Tokenization mechanisms · p. 37
  • Time-series foundation models commonly use encoder-decoder or decoder-only transformer backbones, with relative positional embeddings helping encoder-decoder models handle long sequences.TimeGPT [196] encodes historical temporal dependencies and decodes future values, supporting trends and seasonality across telecom prediction tasks without extra training.Source: 2) Tokenization mechanisms · p. 37, Figure 17 · p. 38
  • Relative positional embeddings, bidirectional attention, and causal attention distinguish the reviewed backbones and determine how historical information is represented for forecasting.Encoder-decoder models let each token attend bidirectionally, whereas decoder-only models restrict attention to preceding tokens.Source: 2) Tokenization mechanisms · pp. 37–38
  • Decoder-only models use causal attention, whose triangular attention matrices improve expressibility and align prediction with historical-to-future temporal ordering.TimesFM [200] predicts extended future sequences in one step, enabling fewer inference steps and rapid responses to channel-state, traffic, and user-location dynamics.Source: 2) Tokenization mechanisms · pp. 37–38
  • Tokenization and architecture choices are essential for pre-training telecom foundation models on mixtures of datasets generated by many devices and users.Reducing input tokens lowers pre-training difficulty on resource-constrained edge devices processing CSI, traffic-load, and network-performance data.Source: 2) Tokenization mechanisms · p. 38

C. Frozen Pre-trained LLM for Prediction

Frozen pre-trained LLM methods adapt general-domain LLMs to telecom prediction without further fine-tuning. They use prompting-based methods with hard or soft prompts, or preprocessing-based methods that reformat numerical values for LLM tokenization.

Prompting-based methods

  • Prompting-based methods adapt time-series inputs through either fixed textual structures or trainable embeddings while keeping the pre-trained LLM frozen.Hard prompts concatenate textual instructions with the input sequence, whereas soft prompts learn embeddings that guide prediction.Source: C. Frozen Pre-trained LLM for Prediction · pp. 38–39

Preprocessing-based methods

  • Preprocessing-based methods transform numerical values into tokenization-compatible textual strings, allowing prediction through standard language-model sampling without relying on specific language structures.LLMTIME introduces digit spacing and decimal-point elimination with rescaling to shape how numerical patterns are tokenized.Source: C. Frozen Pre-trained LLM for Prediction · pp. 39–40

Prompting-based methods

  • Hard prompts convert numerical sequences into natural-language-like contexts, enabling zero-shot prediction and comparable or superior accuracy across various datasets.Input prompts provide historical context and identify the target time step, while output prompts specify the desired prediction value; telecom examples include traffic load, user number, and customer service prediction.Source: C. Frozen Pre-trained LLM for Prediction · p. 39, Table X · p. 40
  • Soft prompts combine textual domain knowledge and in-context examples with patch embeddings, training only a custom embedding layer while the pre-trained LLM parameters remain frozen [47].TIME-LLM uses patch reprogram to encode time-series inputs and concatenates the resulting patch embeddings with prompt embeddings before output projection.Source: C. Frozen Pre-trained LLM for Prediction · p. 39, Figure 19 · p. 39

Preprocessing-based methods

  • Preprocessing-based methods avoid careful prompt design and can generate prediction results instantly from raw network data across telecom prediction tasks.These techniques provide a simple and efficient way to adapt LLMs to varied telecom prediction problems.Source: C. Frozen Pre-trained LLM for Prediction · p. 40

D. Fine-tuned LLM Prediction · E. Multi-modal LLM for Telecom Prediction · F. Discussions and Analyses

The section presents parameter-efficient fine-tuning, multimodal inputs, and practical considerations for LLM-based telecom prediction. It emphasizes efficient deployment, richer prediction contexts, and data and method requirements.

D. Fine-tuned LLM Prediction

  • LoRA freezes pretrained weights and learns low-rank updates, retaining general capabilities while incorporating time-series domain knowledge.LLM4TS [205] applies LoRA to query and key matrices and achieves state-of-the-art performance.Source: D. Fine-tuned LLM Prediction · p. 40
  • LNT adapts layer-normalization scale and shift parameters, and LLM4TS [205] combines LNT with LoRA for time-series prediction.Another strategy freezes attention and feed-forward layers while fully fine-tuning embeddings and applying LNT.Source: D. Fine-tuned LLM Prediction · p. 40

E. Multi-modal LLM for Telecom Prediction

  • Multimodal inputs address limitations of single-modality telecom prediction by incorporating environmental context, historical data, images, and user instructions.For CSI and beamforming, weather maps, building distributions, historical tracks, instant images, and text can jointly describe signal conditions and blockages.Source: E. Multi-modal LLM for Telecom… · p. 41
  • QoE prediction benefits from jointly evaluating users’ textual comments and network metrics, preserving service-specific experiences that attribute extraction may lose.Multimodal LLMs can also generate and predict user experience using comprehension and reasoning capabilities.Source: E. Multi-modal LLM for Telecom… · p. 42

F. Discussions and Analyses

  • Large-scale time-series datasets are prerequisites for pretraining Time-LLMs, but telecom data collection is difficult because of heterogeneous sources, operators, and privacy constraints.Table XI summarizes LLM-based prediction features, inputs, fine-tuning requirements, advantages, and application opportunities.Source: F. Discussions and Analyses · p. 42, Table XI · p. 43
  • Prompting and preprocessing are efficient LLM prediction approaches for short-term telecom tasks requiring instant responses.Prompting leverages zero-shot learning, while converting numerical values into textual strings enables standard language processing capabilities.Source: F. Discussions and Analyses · p. 42

D. Fine-tuned LLM Prediction

E. Multi-modal LLM for Telecom Prediction

VIII. CHALLENGES AND FUTURE DIRECTIONS OF LLM-EMPOWERED TELECOM … 3) Prompt engineering for telecom applications

The section identifies data, deployment, and prompt-design challenges in telecom LLMs, and outlines solutions including comprehensive datasets, hybrid deployment, efficient models, and standardized prompting templates. These directions target effective, practical, and context-aware LLM use across telecom networks.

VIII. CHALLENGES AND FUTURE DIRECTIONS OF LLM-EMPOWERED TELECOM

1) Telecom-domain LLM training:

  • Existing telecom datasets are typically small and task-specific, motivating comprehensive resources spanning documents, specifications, protocols, textbooks, papers, and other network sources.One dataset contains 10,000 telecom questions and answers, approximately 25,000 pages, and 6 million words.Source: 1) Telecom-domain LLM training · p. 43
  • Telecom LLM training should combine general-corpus pre-training with communication-network fine-tuning while balancing model size, performance, energy use, and computational resources.Model compression, knowledge distillation, and specialized hardware can improve efficiency for capacity-limited vehicles, mobile phones, and other devices.Source: 1) Telecom-domain LLM training · pp. 43–44
  • Integrating released telecom datasets for summarization, question answering, and sentence classification can improve customer support, network-issue mitigation, and fraud detection.These datasets are mainly extracted from telecom-related documents and target improved operational efficiency and customer satisfaction.Source: 1) Telecom-domain LLM training · p. 44, Table XII · p. 44

2) Practical LLM deployment in telecom

  • Hybrid deployment across cloud, edge, and user devices can balance scalability, latency, privacy, and generation quality, but coordinating models for diverse tasks requires dedicated designs.Large models such as GPT-4 and LLama3-70b suit high-quality cloud tasks, whereas smaller models support latency-sensitive edge or device tasks.Source: 2) Practical LLM deployment in telecom · p. 44
  • Cloud-only inference can violate latency requirements for wireless applications, with LLM inference times ranging from 0.58 to 90 seconds, while edge servers face resource constraints.Latency-sensitive image and video tasks further incur upload and download delays under limited backhaul capacity.Source: 2) Practical LLM deployment in telecom · p. 44

3) Prompt engineering for telecom applications

  • Telecom prompts must balance specificity and generality while incorporating network context, because the domain’s protocols, technologies, and evolving situational variables complicate concise instruction design.Explicit context cues and conditional generation can improve troubleshooting responses involving logs, performance issues, or security vulnerabilities.Source: 3) Prompt engineering for telecom… · pp. 44–45
  • Standard prompting templates can guide telecom tasks by specifying objectives, control variables, feedback, or prior examples, with researchers and industry experts shaping domain-appropriate standards.Such templates aim to produce precise, pertinent, and impartial responses across telecom applications.Source: 3) Prompt engineering for telecom… · p. 45

B. Future Directions … IX. CONCLUSION

The survey identifies multimodal understanding, automated planning, optimization, enhanced machine learning, practical deployment, hallucination mitigation, retrieval augmentation, and affordable models as key opportunities and challenges for LLM-enabled telecom. It concludes that LLMs could bring artificial general intelligence to telecom through in-context and zero-shot learning, while datasets, cost, reliability, and deployment efficiency remain important concerns.

1) Multi-modal LLMs for telecom:

  • Multimodal LLMs could combine text, images, audio, video, and 3D data to support holographic wireless-environment understanding, blockage prediction, CSI estimation, and video-to-text telecom analysis [30].Applications include mmWave/THz beamforming, traffic-accident reporting [17], and UAV-based descriptions of signal-transmission environments.Source: 1) Multi-modal LLMs for telecom · p. 45

2) LLM-enabled planning in telecom:

  • LLMs require better automated multi-step planning because complex telecom tasks involve sequential decisions, task decomposition, coupled controls, and network scheduling.Potential solutions include structured reasoning, automated decomposition, and simulation-based trial-and-error, while existing prompting methods [14] and often rely on manually decomposed subtasks.Source: 2) LLM-enabled planning in telecom · p. 45

3) LLM for resource allocation and network optimization; 4) LLM-enhanced machine learning for telecom:

5) Real-world implementations of LLM for telecom industry:

  • On-device LLMs and natural-language data interfaces can improve telecom privacy, latency, connectivity resilience, operational efficiency, and employee access to network insights [106],,.Model compression is needed for edge and mobile deployment; one on-device inference model achieved 9.3× faster generation [112], although compression can reduce performance.Source: 5) Real-world implementations of LLM… · p. 46

7) Overcoming hallucination problems in telecom applications:

  • Reducing hallucinations is essential for reliable telecom services, requiring verified data, post-generation verification, external knowledge, real-time fact-checking, and adversarial evaluation.Under specific evaluation conditions, LLMs have shown zero-shot factual-consistency assessment capabilities [227].Source: 7) Overcoming hallucination problems… · p. 46

8) Retrieval augmented-LLM for telecom:

  • Retrieval augmentation can ground telecom LLMs in current external knowledge and improve response relevance, but longer contexts can increase computation, latency, and service degradation.Future work should optimize indexing, search, and query-dependent retrieval to balance context relevance and length.Source: 8) Retrieval augmented-LLM for telecom · pp. 46–47

IX. CONCLUSION

  • The survey concludes that LLMs offer strong potential for artificial general intelligence in telecom through in-context and zero-shot learning, while datasets, cost, and future research challenges require continued study.It reviews LLM fundamentals, applications across generation, classification, optimization, and prediction, and deployment opportunities for researchers and engineers.Source: §IX · p. 47

Related papers

Loading 2405.10825v2…