Large Language Model-based Optimization for Wireless Communication: A Survey
Organizations: Department of Electrical and Electronic Engineering, Hanyang University, Ansan 15588, South Korea · Department of Computer Science, College of Computing, Hanyang University, Ansan 15588, South Korea
Abstract
The growing complexity and dynamism of modern wireless communication frameworks have made real-time optimization essential for efficient resource allocation. Traditional optimization and machine-learning techniques often struggle with generalization, adaptability, stability, and scalability in heterogeneous, time-varying environments. Large language models (LLMs) have recently emerged as powerful tools capable of reasoning over structured prompts and generating context-aware decisions in real time without domain-specific retraining. This survey provides a comprehensive review of LLM-driven optimization techniques applied to wireless communication problems such as power control, bandwidth allocation, task offloading, and intelligent surface configuration. We introduce a functional taxonomy of LLM-enabled wireless resource management strategies and analyze the prompt engineering methodologies employed in recent literature. Furthermore, we develop a generalized analytical framework that categorizes LLM-based optimization architectures, including direct solution generation, closed-loop optimization with external verifiers, and representation encoding within trainable neural pipelines. To demonstrate practical applicability, we present representative case studies on LLM-assisted beam prediction for millimeter-wave systems and adaptive wireless resource allocation. Finally, we discuss emerging hybrid AI approaches for wireless networks.
Figures & tables
| Functional area | Key papers | Pipeline role | Prompting style | LLM role | LLM models used | Baseline methods compared |
| Power control | [ 60 ] | Direct optimizer | Few-shot prompting | Prompt-driven power controller | LLaMA-3-8B-Instruct, GPT-3.5 Turbo, LLaMA-3-70B | DRL |
| [ 63 ] | Objective/constraint interpretation support | Instructional + contextual prompting | Prompt-driven power optimizer | GPT-3.5 | DPC, GenAI agents alone | |
| [ 62 ] | Tool augmented optimization | Instructional + Chain-of-thought + Contextual prompting | Tool-using code generator | GPT-4, GPT-3.5 | Optimal water-filling solution, full knowledge scheme | |
| [ 64 ] | Objective/constraint interpretation support | Chain-of-Thought + Role-based prompting | Intent-processor and expert-selector | GPT-3.5 Turbo | Traditional gate network-enabled MoE approach | |
| [ 65 ] | Objective/constraint interpretation support | Instructional + role-based prompting | Prompt-based policy generator | LLaMA-3.1-8B-Instruct | DDPG with cosine decay noise, DDPG with linear decay noise, Equal power allocation | |
| [ 66 ] | Objective/constraint interpretation support | Instructional + Contextual prompting | Prompt-guided RIS controller | General-purpose LLMs | Random RIS phase configuration, Non-RIS assisted IoV, RIS with fixed strategies |
| Functional area | Key papers | Pipeline role | Prompt style | LLM role | LLM models used | Baseline compared |
| Base station and RIS deployment | [ 76 ] | Direct optimizer | Instructional + iterative + RAG prompting | Prompt-based BSS optimizer, Autonomous BSS agent, Cooperative multiple BSS agents | LLaMA, Bloom, GPT | Particle swarm optimization (PSO), simulated annealing (SA) |
| [ 77 ] | Direct optimizer | Instructional + contextual prompting | Fine-tuned BS position predictor | DistilBERT | CNN-only actor, LLM-only actor, combined CNN+LLM actor | |
| [ 78 ] | Direct optimizer | Instructional + contextual + iterative prompting | Prompt-driven UAV planner | GPT-4 Turbo, GPT-4o-mini | Deep Q network (DQN), optimal theoretical upper bound | |
| Integrated sensing and communication | [ 79 ] | Direct optimizer | Instructional + iterative + contextual prompting | Prompt-guided solver | GPT-3.5 Turbo, GPT-4.0 | NSGAII, MOEAD, RVEA, AGE-MOEA, MODEA |
| Communication standards understanding | [ 83 ] | Objective/constraint interpretation support | Instructional + contextual + RAG prompting | Retrieval-augmented knowledge assistant | Phi-2, GPT-4o, GPT-4o-mini | Base Phi-2, GPT-4o-mini, TeleQnA datasets |
| [ 84 ] | Objective/constraint interpretation support | Instructional + contextual + RAG prompting | Retrieval-augmented knowledge assistant | GPT-4, GPT-3.5, Gemini Pro 1.0 | TSpec-LLM dataset |
| Functional area | Representative works | Objective / goal | Representative constraints | Verifier / evaluation | Main metrics | Timescale |
| Power control | [ 60 , 61 , 62 , 67 ] | Maximize rate, EE, SE, or QoS; minimize transmit power | Power budget, interference, QoS/rate/SINR, RIS/O-RAN limits | Simulation, external rate/SINR evaluation, DRL/search/baseline comparison | Sum rate, EE, power, QoS, convergence | Near-RT / non-RT |
| Bandwidth allocation | [ 66 , 68 , 69 ] | Maximize rate, fairness, reliability, slice utility, or resource efficiency | RB/PRB budget, OFDMA assignment, SINR/QoS, queue state, slice limits | MATLAB/Python feedback, simulation, OAIC/O-RAN testbed | Rate, fairness, delay, reliability, utility | Near-RT / non-RT |
| Mobile edge computing | [ 70 , 71 ] | Minimize service latency; optimize edge–cloud offloading | Server capacity, wireless/backhaul capacity, RBs, LLM delay, task quality | Simulation, reward/score feedback, DRL/GA/exhaustive baselines | Latency, reward, success rate, quality, convergence | Near-RT / non-RT |
| Traffic prediction and analysis | [ 67 , 72 , 73 , 74 ] | Predict traffic states and support proactive resource planning | Historical traffic, spatial–temporal dependence, logs, privacy, data scarcity | Dataset benchmarks, zero-shot tests, Wi-Fi/network-analysis evaluation | MAE, accuracy, throughput, delay, code quality | Offline / non-RT |
| Beam prediction | [ 75 ] | Predict future beam index/direction; maximize beamforming gain | Mobility, mmWave channel variation, codebook, AoD, prediction horizon | DeepMIMO benchmark, LSTM baselines, robustness and ablation tests | Beam gain, accuracy, MSE, robustness | Real-time / near-RT |
| BS/RIS/UAV deployment | [ 76 , 77 , 78 ] | Improve coverage, received power, user rate, QoS, and placement efficiency | Site/location limits, coverage radius, traffic demand, blockage, path loss, IAB backhaul | Real-world data, RAG/agent optimization, Sionna ray tracing, RL baselines | Coverage, cost, received power, rate, iterations | Offline / non-RT |
| Measurement | BP-LLM | Cascaded LSTM |
|---|---|---|
| Avg. latency (ms) | 4.73 | 1.74 |
| Peak memory (MB) | 949.67 | 24.73 |
| Avg. power (W) | 22.55 | 37.41 |
| Energy / inference (J) | 0.137 | 0.054 |
| Metric | FP16 | Q8_0 | Q4_0 |
|---|---|---|---|
| Valid output (%) | 100.0 | 94 | 89 |
| Normalized cost | 12.24 | 11.98 | 11.85 |
| Coverage (%) | 96.68 | 94.31 | 93.50 |
| Latency (ms) | 1685.17 | 1188.02 | 936.68 |
| Energy (J) | 416.75 | 283.32 | 211.14 |
| Peak GPU memory (MB) | 23313.18 | 14929.18 | 10747.18 |
| Mode | Coverage | Latency | Speedup |
| (%) | (ms) | ||
| GPU | 96.13 | 1814.99 | |
| CPU | 96.13 | 38203.16 |
Large Language Model-Guided Optimization
Large Language Model-Guided Optimization uses language models to propose, steer, and refine solutions to hard search and design problems.
See all 654 papers