ACCEPTED FROM OPEN CALL

L L M  W

N: A O   P
E P
Hao Zhou, Chengming Hu, Dun Yuan, Ye Yuan, Di Wu, Xi Chen, Hina Tabassum, and Xue Liu

AbstrAct complex large-scale systems with various knowl-

Recently, large language models (LLMs) have edge domains, that is, signal processing and trans-
been successfully applied to many fields, showing mission, network architecture and design, protocol,
outstanding comprehension and reasoning capa- standards, and so on. Applying general-domain
bilities. Despite their great potential, LLMs usually LLMs directly to domain-specic network tasks may
require dedicated pre-training and fine-tuning for lead to poor performance. Secondly, the devel-
domain-specific applications such as wireless net- opment of LLMs relies on high-quality datasets
works. These adaptations can be extremely demand- for ne-tuning adaptation, while there is a limited
ing for computational resources and datasets, while amount of high-quality networking datasets such as
most network devices have limited computation SPEC5G and Tspec-LLM [1]. Moreover, LLMs are
power, and there are a limited number of high-qual- extremely demanding in terms of computational
ity networking datasets. To this end, this work resources. LLM pre-training and ne-tuning are usu-
explores LLM-enabled wireless networks from the ally implemented on high-performance GPUs such
prompt engineering perspective, that is, designing as NVIDIA A100 and H100, but wireless network
prompts to guide LLMs to generate desired output devices usually have limited computational and
without updating LLM parameters. Compared with storage capacities. LLMs involve a broad range of
other LLM-driven methods, prompt engineering can techniques, such as pre-training, ne-tuning LLMs
better align with the demands of wireless network for domain-specic tasks, retrieval augmented gen-
devices, for example, higher deployment exibility, eration (RAG), prompt engineering, and so on.
rapid response time, and lower requirements on Therefore, it is critical to identify an ecient meth-
computation power. In particular, this work first od to better adapt LLMs to wireless networks.
introduces LLM fundamentals and compares dier- Given the above opportunities and challenges,
ent prompting techniques such as in-context learn- this work introduces prompt engineering, which
ing, chain-of-thought, and self-renement. Then we is regarded as a resource-efficient and flexible
propose two novel prompting schemes for network approach to using LLMs with fast implementation
applications: iterative prompting for network opti- speed [4]. These advantages will help to overcome
mization, and self-refined prompting for network the above LLM application challenges, for exam-
prediction. The case studies show that the proposed ple, deployment difficulties and requirements for
schemes can achieve comparable performance computational resources. In particular, prompting
as conventional machine learning techniques, and refers to designing input prompts to guide pre-
our proposed prompting-based methods avoid the trained LLMs to generate desired outputs. It takes
complexity of dedicated model training and ne-tun- advantage of the inherent inference capabilities of
ing, which is one of the key bottlenecks of existing pre-trained LLMs, and avoids the need for back-
machine learning techniques. ward passes and gradient updates. Therefore,
prompt engineering has several key features:
IntroductIon • Resource-efficient: Prompting only needs for-
As a sub-field of generative AI, large language ward passing of the model, and no need to
models (LLMs) have received considerable interest store all intermediate activations for back-prop-
from industry and academia [1]. The advancement agation. Such a resource-efficient approach
of generative AI and LLMs also provides promising may mitigate the computational burden of net-
opportunities for 6G networks, including strong work servers and devices.
reasoning and planning capabilities, multi-modal • Higher flexibility: Prompting-based methods
understanding for 6G sensing, semantic communi- can quickly adapt to various tasks by crafting
cation [2], integrated satellite-aerial-terrestrial net- the corresponding textual demonstrations and
works [3], and so on. Despite the great potential, queries without extra coding steps[5]. It indi-
integrating LLMs into wireless networks still faces cates an ecient method to customize LLMs to
several challenges. Firstly, wireless networks are address a wide range of network tasks.

Hao Zhou, Chengming Hu, Dun Yuan, Ye Yuan, Di Wu, Xi Chen, and Xue Liu are with McGill University, Canada;
Digital Object Identifier: 10.1109/MWC.001.2400384 Hina Tabassum is with York University, Canada.

98 1536-1284/25/$25.00 © 2025 IEEE IEEE Wireless Communications • August 2025

• Fast implementation: Prompt engineering relies LLMs to develop fundamental capabilities such as Instead of updating
on the inference capabilities of LLMs, and it comprehension, reasoning, and instruction follow- model parameters,
avoids the time cost of updating LLM parame- ing. Then, these capabilities are the foundation for prompt engineering
ters. Therefore, the low response time can bet- developing other LLM-inspired techniques such as utilizes LLM’s inference
ter handle low-latency network services. prompt engineering and planning, for example,
Finally, note that although prompting LLMs can prompt engineering relies on the inherent capa- capabilities, designing
significantly enhance the performance, it is not bilities of LLMs. However, it requires considerable specic input prompts
a standalone solution for LLM usage[6]. For computational resources such as NVIDIA H100- to guide the generation
instance, combining prompt engineering with 80G or A100-80G for implementation, and the of pre-trained LLMs.
fine-tuning may create a robust framework with training process may take weeks and months. Importantly, prompt
both prompting flexibility and model specializa- Given the training costs, pre-training a wire-
tion. However, this work focuses on prompting less-specific LLM from scratch can be extremely engineering avoids the
techniques to investigate their full potential. time and energy-consuming. By contrast, ne-tun- need for backward
LLMs have been discussed in several existing stud- ing is a more affordable approach for adapting passes and gradient
ies, but they mainly focus on system-level discussions LLMs to network domains. Given pre-trained gen- updates of LLMs, and
and module designs, for example, edge intelligence eral-domain LLMs, fine-tuning indicates adapting therefore the time costs
[7] and grounding and alignment[8]. Prompting engi- LLMs to specific tasks by partially updating the
neering is also used for AI-generated everything ser- model weights to fit smaller and domain-specific are signicantly lower
vices in [9], and our previous work also investigates datasets such as 3GPP, 5G, and O-RAN standards. than ne-tuning.
LLM-enabled power control and trac prediction in The ne-tuning costs are related to the LLM model
[10, 11]. However, this work is dierent from exist- size and fine-tuning methods. For instance, fully
ing studies by systematically exploring prompt engi- fine-tuning a 7B LLM model using 8 A100 GPUs
neering and wireless network applications, providing may take 100–200 GPU hours for a 100M token
detailed prompt designs and specific case studies. dataset. For large-scale datasets in wireless fields,
The main contributions are as below. for example, SPEC5G dataset with 134M words
Firstly, we present in-depth analyses of LLM and 3GPP dataset Tspec-LLM with 534M words
fundamentals and the feasibility of network appli- [10], the costs on computational resources may
cations such as pre-training, ne-tuning, RAG, and be larger. Similarly, RAG is another promising
prompt engineering. Each technique is introduced direction for practical LLM applications for wire-
in terms of computational resources and time less networks. It integrates LLMs with an external
costs, potential diculties, and wireless network knowledge base, and then the LLM can retrieve
applications. Then, we provide an overview of a related information from the external knowledge
variety of prompting techniques and discuss their source during the inference. Since LLMs have
applications to wireless networks, such as in-con- access to the most relevant and up-to-date informa-
text learning, chain-of-thought, prompt-based tion, for example, novel network standards, archi-
planning, and self-renement. tecture and signal transmission techniques, RAG
We propose two novel prompting techniques, can signicantly improve the generation quality.
namely iterative prompting and self-refined Instead of updating model parameters, prompt
prompting, aiming to address network optimiza- engineering utilizes LLM’s inference capabilities,
tion and prediction problems, respectively. Spe- designing specic input prompts to guide the gen-
cically, iterative prompting learns from previous eration of pre-trained LLMs. Importantly, prompt
experiences to improve LLM’s performance on engineering avoids the need for backward pass-
target tasks iteratively, which is suitable for han- es and gradient updates of LLMs, and therefore
dling network optimization problems. By contrast, the time costs are signicantly lower than ne-tun-
self-refined prompting can correct their out- ing. Due to the low costs, it can quickly adapt to
puts through iterative feedback and refinement dynamic network environments, handling various
prompts, aiming to address network prediction network tasks with much lower response time. In
tasks. These schemes rely on LLM’s inference addition, prompting engineering also aligns well
capabilities, and such resource-efficient tech- with several crucial wireless network features.
niques align well with the limited computational Firstly, many network devices are computation-
capacities of many network devices. In addition, al and energy resource-constrained, and prompt
our proposed algorithms can be easily general- engineering has low requirements for computa-
ized to various wireless network applications, and tional resources, which may mitigate the energy
the case studies show that they achieve satisfac- consumption and computational burden on net-
tory performance in network power control and work devices. Secondly, prompt engineering allows
trac prediction problems. human language-based instructions, for example,
network-specic task descriptions, questions, and
LLM FundAMentALs towArd solutions. By integrating natural language, network
intelligence will become more accessible, especial-
wIreLess network AppLIcAtIons ly for operators with minimal professional AI knowl-
This section introduces various approaches for edge. In addition, prompt engineering has high
integrating LLMs to wireless networks, including design flexibility. With proper textual templates,
pre-training, ne-tuning, RAG, and prompt engineer- prompts can be eciently designed using human
ing. Specically, Table 1 presents these techniques in language and applied to various network tasks,
terms of their working principle, advantages, poten- which also have minimum requirements on math-
tial difficulties, computational resources and time ematical equations. Moreover, prompting can be
costs, and possible wireless network applications. easily combined with other techniques, for exam-
Firstly, pre-training is a foundational step for ple, ne-tuning LLMs on network-specic datasets,
LLM development, which includes data collection, and then prompting the model to further improve
ltering, and model training. Pre-training enables the generated content.

IEEE Wireless Communications • August 2025 99

 Computational resources Possible wireless network
Approaches Key principles Advantages Potential difficulties
and Time costs applications

Pre-training
1) Training LLMs from High Thermal Design A network-specic LLM
enables LLMs with Pre-training is extremely
scratch on hundreds of Power GPUs such NVIDIA indicates great benets,
comprehension, resource-intensive,
billions of tokens for the H100-80G or A100-80G but training network LLMs
Pre-training reasoning, and requiring long training
next token prediction. are required, e.g., training from scratch may be
LLMs instruction following times, and substantial
2) Pre-training is a Llama3.1-405B took inappropriate due to the
capabilities, which energy consumption
foundational step of LLM approximately 30.84M computational resource
can address various and GPUs or TPUs.
development and usage. GPU hours. requirements.
downstream tasks.

Fine-tuning refers to The cost may vary

Fine-tuning signicantly 1) Fine-tuning has to Fine-tuning is a realistic
adapting a pre-trained between model sizes and
reduces time and consider the risk of approach to apply LLMs
LLM to specic tasks by ne-tuning methods. If full
computational cost overtting and losing to wireless networks,
partially or completely ne-tuning a 7B parameter
Fine-tuning compared to pre- generalization. adapting a general-domain
updating the model model with 8 A100-80G
training, producing 2) Labelled data may be LLM to specic tasks, e.g.,
weights or introducing GPUs, it could be 100-200
LLM models tailored to needed with additional network troubleshooting
additional adapter GPU hours on a 100M
specic tasks. human labour. and conguration.
modules. token dataset.

1) RAG is a more
1) RAG involves extra 1) The cost depends on
RAG combines LLMs practical approach for
integration complexity, the retrieval method and
with an external RAG can improve implementing LLMs,
as it requires ecient the size of the knowledge
Retrieval knowledge base, LLM’s response especially when existing
retrieval algorithms. base. Inference may take
augmented enabling the model accuracy by leveraging network-specic datasets
2) RAG highly depends seconds to minutes per
generation to retrieve relevant external knowledge are available.
on a well-structured query.
(RAG) information during sources, generating 2) It ensures the LLM has
knowledge base, 2) Requires a balance
inference and provide more up-to-date results. access to the most relevant
requiring dedicated between retrieval speed
more accurate responses. and up-to-date network
creation. and accuracy.
data.

1) Without the need

1) Prompt engineering 1) It requires expertise
Prompt engineering for backward passes 1) Prompt engineering
has low computational in crafting eective
refers to crafting specic and gradient updates, enables rapid decision-
costs and quick prompts.
inputs to guide a pre- it requires minimal making and responses in
Prompt implementation. 2) It is also limited by
trained LLM to generate computational resources. dynamic wireless network
engineering 2) It can be easily the inherent capabilities
desired outputs, without 2) Time costs are environments.
adapted to dierent and context window
any model retraining or signicantly lower 2) It aligns well with many
tasks without additional size of the pre-trained
tuning. compared to other network tasks.
training. LLMs.
methods.

Note that the objective of this table is to introduce the features of each technique, and they serve dierent purposes in LLM development and application.
These approaches can be combined to achieve better performance, e.g., ne-tuning network LLMs and then prompting.
TABLE 1. Summary of various LLM usage approaches for wireless networks.

On the other hand, prompt engineering also text learning, serving as critical references for LLMs
requires considerable expertise in crafting the to learn from. Therefore, they should be careful-
prompts, which will directly affect the quality of ly selected, formatted, and designed, which may
generated content. Although prompt engineer- require professional understanding and knowl-
ing can signicantly improve LLM’s performance, edge. In-context learning is a promising approach
it relies on the model’s inherent capabilities, for to applying LLMs to wireless networks, which can
example, achieved by pre-training or ne-tuning. take advantage of previous network solutions as
Therefore, given the opportunities and challenges, demonstrations to address unseen network issues.
it is crucial to explore the integration of prompt- Meanwhile, wireless networks are complicated
ing techniques and wireless networks, showing a systems, and one task may include several logic and
promising way to utilize generative AI. reasoning steps. For instance, many network opti-
mization problems include several elements, e.g.,
proMpt engIneerIng For wIreLess networks base stations, transmission power, bandwidth, and
This section presents several prompting tech- network users. Using simple input-output pairs as
niques and their network applications, that is, demonstrations cannot provide the logical reasons
in-context learning, chain-of-thought, prompt- behind these examples, and LLMs may have dicul-
based planning, and self-renement methods. ty in learning. To this end, chain-of-thought is pro-
Firstly, as shown in Fig. 1, in-context learning posed as an advanced strategy to enhance LLM’s
is a fundamental and crucial capability of LLMs. It performance in complex reasoning tasks. Specifi-
refers to the process of learning from formatted cally, chain-of-thought will explain each logic step in
natural language instructions and demonstrations, the prompt demonstrations: “Let’s think step by step.
and then improving the performance on target Given the current base station transmission power
tasks. Given contextual demonstration prompts, {BS_power}, using the Shannon capacity equation,
LLMs can address various downstream tasks by the transmission data rate is {user_rate}. Then, we
using existing knowledge from pre-trained data. compare the achieved rate with the target data rate,
Demonstrations are of great importance for in-con- and {user_rate} is lower. Therefore, we may need

100 IEEE Wireless Communications • August 2025

FIGURE 1. Illustration of dierent prompting techniques.

to increase the transmission power.” This example tion proposes a novel iterative prompting scheme,
includes two steps, “using Shannon capacity equa- in which we prompt LLMs iteratively to improve
tion” and “compare the achieved rate with the target the performance on target network tasks.
data rate,” and then the nal decision is to “increase
the transmission power.” With these two-step expla- nAturAL LAnguAge-bAsed tAsk descrIptIon
nations, LLMs can easily capture the relationship Firstly, as shown in Fig. 2, the target task is
between transmission power and target data rate, described by formatted natural language, includ-
and generate better replies using the logical chain. ing task goals, definitions, and extra rules, pro-
In-context learning and chain-of-thought viding fundamental task information to the LLM.
are mainly designed for single tasks, which may For instance, the task goal describes the problem
have difficulty tackling more complicated tasks, as a “decision-making task,” and it introduces the
for example, collecting information from multiple set of control variables as “candidate decisions.”
sources and producing multi-step reasoning to Then, it highlights the importance of “environment
generate a comprehensive answer. Many complex states,” indicating that the LLM has to consider
network tasks need to be broken down into more specic environment variables for decision-mak-
manageable sub-tasks. For example, the network ing. After that, it introduces the current target task
conguration usually involves multiple interrelated by “here is a new case to solve, and the current
network elements, and network project develop- environment state is….” Moreover, the nal rules
ment may consist of a series of coding and testing specify that “select from {Decision_set} based on
tasks. In this case, prompt-based planning aims to above examples.”
generate a group of sub-tasks, which will be exe- With these formatted task descriptions, the
cuted one by one. This planning capability is cru- LLM can capture the key elements of the deci-
cial for handling many large-scale network tasks. sion-making task, focusing on hidden patterns
In addition, LLMs may generate incorrect between environment states and corresponding
answers at the initial attempts, and external feed- decision variables. Such a design can also lower
back is essential to improve its replies. As shown the understanding difficulties caused by profes-
in Fig. 1, self-refinement allows LLMs to self-im- sional network knowledge, since general domain
prove their previous outputs by generating feed- LLMs such as GPTs and Llama are not specically
back and refinement prompts. For instance, the pre-trained for network applications.
LLM’s initial transmission power decision may be
incorrect, and we can use the same LLM to eval- IterAtIve proMptIng scheMe
uate and provide feedback, and then feed the As illustrated in Fig. 2, the above task description
feedback and renement to the LLM, improving will become input prompts for the LLM at network
the initial power control decisions. In particular, edge cloud, and the LLM selects a specific Deci-
it indicates that LLMs can automatically improve sion_n from the set of candidate decisions {Deci-
their previous outputs without human interven- sion_set}. Then, the network management decisions
tion. Self-renement can signicantly save human will be sent to the wireless network environment for
eort in improving LLM’s initial replies, especially implementation such as resource allocation, user
considering the complexity of network tasks. association and transmission power level decisions.
Finally, Table 2 compares the above techniques The network operation results be collected as a
in terms of main features, advantages, potential new example, including environment states, select-
issues, and wireless network applications. Note ed decisions, and rewards that are related to optimi-
that these methods can be combined to address zation objectives and network metrics.
complex tasks comprehensively. For instance, After that, the examples from the network out-
using self-renement in prompt-based planning to put, that is, previous explorations, are collected in
save eort in the evaluation phase, and integrating an experience pool, recording all previous exam-
chain-of-thought to improve the multi-step logic ples. The examples in this pool will be further evalu-
reasoning in sub-task decomposition. ated to choose specic examples well-suited for the
target task. For instance, a straightforward strategy
IterAtIve proMptIng For is to select previous examples with similar environ-
ment states as current states, for example, similar
wIreLess network optIMIzAtIon user numbers and channel conditions. Then the pre-
Optimization techniques are of great importance vious decisions and rewards become useful refer-
for wireless network management, and this sec- ences for LLM decision-making. However, note that

IEEE Wireless Communications • August 2025 101

 Prompting
Key principles Advantages Potential issues Wireless network applications
techniques

LLMs can be generalized to

In-context learning empowers In-context learning is a promising
perform various learning tasks
LLMs to learn from task-specic Demonstrations greatly impact in- approach to applying LLMs to
by utilizing existing knowledge
In-context instruction and demonstrations. context learning, which should be wireless networks, using previously
from pre-training data and
learning These prompts are designed carefully designed, e.g., selection, accumulated solutions as input
acquiring new task-solving
with task-specic examples that format, and order of examples. prompts to address new network
strategies from contextual
are comprehensible to LLMs. issues.
demonstration prompts.

1) Chain-of-thought is an Chain-of-thought can make LLMs

LLMs can solve complex Chain-of-thought prompting faces
advanced strategy to improve better adapt to complicated
problems more transparently issues like incorrect reasoning and
LLM’s capabilities in complex wireless environments that
Chain-of- and logically through chain- instability. Their inherent instability
reasoning. require step-by-step reasoning
thought of-thought demonstration may provide incorrect and
2) It can guide LLMs to better and planning, e.g., project
prompts, e.g., “Let’s think step misleading intermediate reasoning
understand the logical steps code generation with multi-step
by step”. steps.
from questions to answers. scheduling.

Prompt-based planning aims LLM-based planning requires Wireless networks are complicated
LLMs can make complex tasks
to address more complicated dedicated analyses to decompose systems and LLM-enabled
Prompt- more manageable by using
scenarios such as multi-hop a complicated task into multiple automated task decomposition
based their planning capabilities,
question answering. It breaks sub-tasks. Zero-shot and can be extremely useful for long-
planning lowering the overall
complex tasks down into automated task decomposition is term network management and
diculties.
manageable sub-tasks. still the main diculty. project development.

Through iterative feedback All historical outputs are iteratively

Self-renement will signicantly
Self-renement allows them to generation and prediction appended as the inputs of LLMs,
save human eorts for network
Self- correct their outputs through renement, LLMs can self- and it may lead to increasingly
tasks, since LLMs can constantly
renement iterative feedback and rening enhance their outputs with lengthy demonstration prompts
improve the previous network
demonstration prompts. strong eectiveness and and unavoidable resource
policy automatically.
scalable capabilities. consumption.
TABLE 2. Summary of dierent prompting techniques for wireless networks.

many wireless environment states are continuous

variables, for example, channel state information or
seLF-reFIned proMptIng For
user-base station distances, and therefore it is unlike- network predIctIon
ly to find examples with exactly the same states. Prediction tasks are critical in wireless networks, using
Therefore, the evaluation and selection of examples historical network data to predict future trends and
are crucial in the proposed scheme. behaviors such as network trac, user demand, chan-
Finally, the selected examples, that is, recom- nel states, and device status. To showcase the poten-
mended and inadvisable demonstrations, are inte- tial of prompting techniques, this section presents a
grated into the task description, providing useful self-refined prompting method for network predic-
references for LLM decision-making. Figure 2 has tion problems, that is, network trac prediction.
presented steps 1 to 6 of the proposed iterative
prompting scheme, and it can iteratively explore predIctIon tAsk descrIptIon
new decisions and accumulate new experiences. A crucial step in using LLMs for prediction prob-
Therefore, the LLM can constantly learn from pre- lems is to convert numerical values into natural
vious explorations, and then improve its decisions language sentences. Figure 3 considers network
accordingly. Such an iterative prompting method trac prediction as an example, and illustrates the
can rapidly adapt to dynamic wireless environments. demonstration prompt, data prompt, and query
Figure 2 applies an exploration-exploitation prompt. In particular, the trac prediction task is
strategy as in reinforcement learning (RL). How- framed in a question-answer format, adhering to
ever, iterative prompting has several advantages a template-based description. The demonstration
over RL approaches: prompt is first incorporated to guide the LLM in
• No model parameter updating and ne-tuning making trac predictions by leveraging contextu-
is required, which has much lower complexity al task descriptions for next-day trac prediction.
than existing RL algorithms such as deep rein- Then, the LLM predicts future trac by respond-
forcement learning (DRL). ing to questions in the query prompt, based on
• It utilizes human natural language for network historical trac data provided in the data prompt,
management. With minimum requirements that is, previous hourly network trac.
on mathematical knowledge and equations, it
makes network optimization more accessible FeedbAck generAtIon And predIctIon reFIneMent
than using RL. Initially, LLMs may generate incorrect predictions,
• LLMs can provide reasonable explanations for and this subsection utilizes self-rened prompting
their decisions, and help humans understand to incorporate feedback demonstrations to pro-
complicated network systems, which is beyond vide comprehensive and valuable insights, aiming
the capabilities of RL techniques. to improve the initial results [11]. As shown in Fig.
Finally, note that our proposed schemes can also 3, the feedback prompt is designed to encompass
be generalized to many other fundamental net- prediction performance, prediction format and
work optimization tasks, providing a novel LLM- completeness, and prediction method, which are
based solution for network control. shown as follows:

102 IEEE Wireless Communications • August 2025

FIGURE 2. Iterative prompting schemes for wireless network optimization.

• Overall performance: The same LLM evaluates prompting method designs task-specific demon-
the overall prediction performance, for exam- stration, feedback, and renement prompts, allow-
ple, Mean Absolution Error (MAE) between ing the LLM to iteratively enhance its predictions
ground truth and predictions. by adhering to the instructions in these prompts,
• Periodical performance: Considering that without requiring additional model ne-tuning or
time-series data can be represented through training. Compared to the dedicated training of
sine and cosine functions in the real domain long short-term memory (LSTM) networks, the
[12], the same LLM is tasked with projecting proposed self-rened prompting method is more
both ground truth and predicted traffic onto computationally ecient with the inference abili-
these functions, respectively. By matching the ties, particularly advantageous for on-device learn-
projection of ground truth, predictions are ing on resource-constrained devices in real-world
expected to capture the periodic nature of wireless networks. Given that iteratively append-
ground truth accurately, such as fluctuations ing all historical predictions can increase the
during peak and o-peak trac periods. length of demonstration prompts and potentially
• Format and completeness: The predictions lead to higher resource consumption, a future
should match the format of ground truth as direction for the self-refined prompting method
well as be complete for each timestamp, for will focus on effectively compressing extensive
example, 24 values for daily trac prediction. demonstration prompts, thereby further enhanc-
• Prediction method: The prediction method is ing resource eciency.
summarized in the feedback demonstration
prompt, which calls for adopting more advanced cAse studIes
and accurate methods to enhance performance
further. To avoid additional computational costs, sIMuLAtIon settIngs
while the LLM may suggest developing a neural This section presents two case studies on network
network as a potential improvement method, optimization and prediction problems.
this recommendation is disregarded. Network Optimization: This case study
Given the above instructions, the refinement involves a base station power control problem a
demonstration prompts aim to provide spe- fundamental optimization task in wireless network
cific actionable steps associated with feedback elds. The considered problem formulations aim
demonstration prompts. This allows the same to minimize the base station power consumption
LLM to self-rene previous predictions by adher- and meanwhile maintain an average data rate
ing to the detailed feedback outlined in feedback threshold for users.1 We consider three adjacent
demonstration prompts. Note that the process of base stations and the associated user numbers
feedback generation and prediction refinement dynamically change from 5 to 15 each, the aver-
is iteratively conducted until the prediction per- age data rate threshold is 1.5 Mb/s/per user, and
formance converges, whereas the inference is the channel gain applies 3GPP urban network
completed without engaging in the feedback gen- models. This case study includes:
eration and prediction renement. Additionally, to • Iterative prompting: We consider 2 LLM mod-
prevent the repetition of previous incorrect pre- els: Llama3-7b-instruct as a small-scale model,
dictions, the history of predicted trac, feedback, and Llama3-70b-instruct as a large-scale model.
and refinement demonstration prompts are also We deploy the proposed iterative prompting
included as inputs to the LLM at each iteration. technique as introduced above, and in Fig. 2,
In summary, LLMs excel at recognizing and in which the LLM will explore the environment,
performing new tasks through prompt engineer- accumulate experience, and learn iteratively.
ing techniques that utilize contextual information, • DRL baseline: DRL is included as a baseline,
including task descriptions and demonstration which has been widely used to address network
examples. Specically, the proposed self-rened optimization problems after dedicated model

IEEE Wireless Communications • August 2025 103

FIGURE 3. Self-rened prompting for network traffic prediction. As an example, by following prediction tasks framed in a question-answer format, the LLM
is guided to predict hourly traffic prediction for 2013-11-08, using hourly historical traffic from 2013-11-07. Due to space constraints, we show traffic for
selected hours only, while hourly traffic is fully utilized during interactions with the LLM. e specic feedback related to initial predictions is incorporat-
ed into the feedback prompt. e LLM further renes its predictions by following the actionable steps outlined in the renement prompt. e process of
feedback generation and prediction renement is iteratively conducted until the prediction performance converges, whereas the inference is completed
without engaging in the feedback generation and prediction renement.

training. We apply a classic deep Q-learning sIMuLAtIon resuLts And AnALyses

algorithm, in which the neural networks have
3 layers, the experience pool size is 10000, the Network Optimization Case Study: Figure
batch size is 64, and the learning rate is 0.005. 4a and b present the optimization task results.
Network Prediction: The evaluation dataset is a Specifically, one can observe that the average
publicly available network trac dataset from the power consumption is constantly improved with
city of Milan [14], recording the time of user inter- the increasing number of episodes in Fig. 4a. It
actions and the base stations managing these inter- indicates that the proposed scheme from above
actions. The original grid of 100¥ 100 base stations enables LLMs to learn from previous examples
is reorganized into 25¥ 25=625 aggregated base and experience, and then improve its perfor-
stations. Each aggregated base station covers an mance on target tasks iteratively. Figure 4a also
area of approximately 1 km¥ 1km, including 8,923 shows that LLMs present comparable perfor-
samples. For the evaluation, we randomly select mance as the DRL algorithm, which is achieved
one base station, partitioning the dataset such that without model training or fine-tuning as in con-
70 percent is used for training, 10 percent for val- ventional ML algorithms. Meanwhile, Fig. 4b
idation, and the remaining 20 percent for testing. compares the service quality of Llama3-7b, Lla-
The simulation algorithms include: ma3-70b, and DRL, which means the probability
• Self-refinement: Our proposed self-refined of violating the preset average data rate thresh-
prompting utilizes GPT-4 as the foundation old of 1.5 Mb/s/per user. It highlights that LLMs
model for one-day-ahead traffic prediction can maintain high service quality at the beginning
using historical traffic from the previous day. of the training process, while the DRL algorithm
It follows the proposed self-rened prompting has a lower service quality in the initial explora-
techniques as dened previously, and in Fig. 3. tion phase. Figure 4b reveals that LLMs have a
• LLM-related baselines: We include GPT-4 and higher learning eciency than the DRL algorithm.
GPT-3.5 without self-renement as two LLM-re- It can be explained by LLM’s few-shot learning
lated baselines, better demonstrating the capa- capabilities, in which LLMs can quickly learn from
bilities of self-renement. demonstrations and task descriptions to improve
• Conventional baselines: Autoregressive inte- decisions. By contrast, the DRL approach relies
grated moving average (ARIMA) and LSTM as on exploration and value updating to improve the
two conventional baselines. Through extensive policy gradually, and the low sampling eciency
experiments, the optimal parameters for the prevents the learning eciency.
ARIMA baseline model are determined includ- Network Prediction Case Study: Figure 4c com-
ing autoregressive (p = 2), dierences (d = 1), pares the average MAE and MSE metrics of vari-
and moving average (q = 2) components. We ous techniques. It is observed that the self-refined
deploy LSTM as an optimal baseline, since it prompting method signicantly outperforms ARIMA,
has been specically trained on the target net- GPT-3.5, and GPT-4, demonstrating the capabilities
1 Detailed equations can work trac dataset. By contrast, Self-renement of our proposed self-refined prediction method.
be found in [13] (equation and LLM-related baselines have no previous The self-refined prompting method exhibits nota-
(1.12) in Section 1.3.5). knowledge on the target task. ble improvements over GPT-3.5 and GPT-4, with

104 IEEE Wireless Communications • August 2025

 a) b)

c) d)

FIGURE 4. Performance comparison in network optimization and traffic prediction case studies: a) Optimization Case
Study: improved power consumption of Llamas; b) Optimization Case Study: service quality comparison with training;c)
Prediction Case Study: average MAE and MSE comparison. e self-rene method achieves comparable performance as
LSTM as a few-shot predictor, while LSTM has been delicately trained on the datase; d) Prediction Case Study: predicted
traffic comparison in 24 hours.

MAE reductions of 29.72 percent and 17.09 per- LLMs also show higher learning efficiency than
cent, which underscores more powerful generaliza- existing techniques, and they allow human lan-
tion capability on unseen datasets through iterative guage-based input, showing the possibility of nat-
feedback generation and prediction renement pro- ural language-based future network management.
cesses. GPT-3.5 shows a worse performance due to
its outdated architecture and designs, which have reAL-worLd chALLenges And LIMItAtIons
been observed in many existing studies. Although This section will further discuss the limitations
LSTM serves as an optimal baseline with superior and real-world challenges of prompt engineering
performance, it is important to note that LSTM is applications.
specifically trained on the target dataset, incurring Firstly, prompt engineering is inherently limited
additional computational costs, which presents signif- by the context window size of LLMs, which restricts
icant challenges in scenarios involving resource-con- the capabilities to process tasks requiring extensive
strained devices within wireless networks. Given the contextual information beyond the model’s input
dynamic evolution of traffic distribution in non-sta- and memory constraints. Recent advancements
tionary wireless networks, a well-trained LSTM may such as Llama3 with a context window size of
overt the historical trac used during the pre-train- 128k tokens, signicantly mitigate this limitation.
ing, resulting in degraded performance on new, Secondly, the eectiveness of prompt engineer-
unseen trac over time. Furthermore, similar results ing relies heavily on the selection of demonstrations
can be observed in Fig. 4d, which illustrates the one- shown to the model or the order of demonstrations.
day ground truth trac and the predicted trac of Poorly chosen or suboptimal examples can result
various methods. The self-rened prediction method in biased outputs. Our work considers iterative
more closely matches the ground truth trac, indi- prompting to select better demonstrations, but pro-
cating strong scalability of predictions across various fessional experiences are still needed in this process.
base stations. By contrast, GPT-4 shows an obvious Moreover, crafting eective prompts for com-
mismatch from 9:00 to 15:00. plicated tasks in network environments usually
In summary, the simulation results in Fig. 4 requires trial-and-error experimentation. This iter-
demonstrate that LLMs have great potential in ative process is inecient, especially when results
handling network management, that is, optimiza- are inconsistent or dicult to reproduce. Without
tion and prediction tasks. They present satisfied clear guidelines, engineers must rely on intuition
performance as conventional machine learning or extensive testing to achieve desired outcomes.
algorithms such as DRL and LSTM, while avoiding This can slow down the development for novel
extra model training and fine-tuning complexity. and complex use cases as in 6G.

IEEE Wireless Communications • August 2025 105

Although LSTM serves Data, vol. 2, no. 1, 2015, pp. 1–15.
In addition, in real-world applications, prompt [15] H. Du et al., “Spear or Shield: Leveraging Generative AI
as an optimal baseline engineering may introduce significant security to Tackle Security Threats of Intelligent Network Services,”
with superior perfor- concerns, as it is potentially vulnerable to attacks arXiv:2306.02384, 2023.
such as prompt injection and leakage. Such
mance, it is important
attacks are particularly concerning in sensitive bIogrAphIes
to note that LSTM Hao ZHou is a Postdoc researcher at the School of Computer
domains like wireless communications, where Science, McGill University. He received PhD degree from Univer-
is specically trained malicious actors could manipulate the model or sity of Ottawa in 2023. His research focuses on the intersection
on the target dataset, extract encoded sensitive data [15]. between AI/ML and networked systems, especially for reinforce-
ment learning and large language models. He received the Best
incurring additional Finally, note that prompt engineering is not a Paper Award from IEEE ComSoc CSIM TC and 2023 IEEE ICC, and
computational costs, standalone solution but one part of pipelines of LLM the best Doctoral Thesis Award from University of Ottawa.
and generative AI techniques [5]. Prompt engineer-
which presents sig- ing is constrained by the inherent limitations of the CHengming Hu is currently a Ph.D. candidate in the School of Com-
puter Science at McGill University. Prior to joining McGill, he received
nicant challenges in used LLMs. If the LLM lacks sucient knowledge or the M.Sc. degree from the Concordia Institute for Information Sys-
scenarios involving context, prompting is unlikely to produce accurate tems Engineering at Concordia University in 2019. His research inter-
results. Therefore, when prompt engineering strug- ests lie at the intersection of computational intelligence techniques
resource-constrained and networked systems, with a particular focus on smart grids, com-
devices within wireless gles with highly specic tasks, it may require exter- munication systems, and cyber-physical systems. His work has been
nal support such as fine-tuning. This dependency published in prestigious venues spanning a range of research topics,
networks. Given the highlights that prompt engineering can be actively including ICLR, IEEE TSG, and IEEE COMST, among others.
dynamic evolution of combined with other techniques such as pre-training Dun Yuan is a Ph.D. candidate at McGill University, working under
traffic distribution in and ne-tuning to make the most of LLM potential. the supervision of Prof. Xue (Steve) Liu in the Cyber-Physical Intel-
non-stationary wireless ligence Lab. His research focuses on integrating reinforcement

networks, a well-trained concLusIon learning (RL) and large language models (LLMs) into the domain of
telecommunication. He also collaborates with the Crypto-Metaverse-
The progress of generative AI and LLMs brings Blockchain-Cloud (CMBC) research group, exploring innovative proj-
LSTM may overt the promising opportunities for next-generation wire- ects in Web 3.0, decentralized storage, and distributed computing
historical traffic used less networks. This work provides a comprehen- systems. He hold a Bachelor’s degree from Xi’an Jiaotong University
and a Master’s degree from Imperial College London.
during the pre-training, sive overview of prompt engineering for network
resulting in degraded applications. It presents in-depth analyses of dif- Ye Yuan is a Ph.D. student in Computer Science at Mila – Que-
performance on new, ferent prompting techniques such as in-context bec AI Institute and McGill University. His research focuses on
learning, chain-of-thought, and self-renement. In score-based generative models with applications on addressing
unseen traffic over addition, this work proposes two novel prompt-
out-of-distribution challenges in oline black-box optimization, as
well as developing foundational knowledge models using diffu-
time. ing schemes for network optimization and pre- sion-based and autoregressive language models. Ye has published
diction problems, and the case studies show that papers in leading venues such as NeurIPS, EMNLP, IEEE COMST,
the proposed methods can achieve satisfactory IEEE WCL. He is a recipient of the BMO Responsible AI Senior
Scholar award and holds a Bachelor of Science in Honours Com-
performance. In the future, we will explore more puter Science from McGill University.
complicated and advanced network tasks such as
recongurable intelligence surfaces and integrat- Di Wu is a research scientist at Canada AI Inc and an Adjunct Profes-
sor at McGill University. He was a senior sta research scientist, a team
ed sensing and communication, uncovering the leader at Samsung AI Center Montreal. He did postdoctoral research
full potential of LLMs for future wireless networks. at Montreal MILA and Stanford University. He received his Ph.D. and
M.Sc. from McGill University in 2018 and Peking University in 2013,
reFerences respectively. Di also holds Bachelor’s degrees in microelectronics and
economics. His research interests mainly lie in reinforcement learning,
[1] H. Zhou et al., “Large Language Model (LLM) for Telecommuni-
cations: A Comprehensive Survey on Principles, Key Techniques, transfer learning, meta-Learning, and multitask Learning. He is also
and Opportunities,” arXiv preprint arXiv:2405.10825, 2024. interested in leveraging such algorithms to improve real-world systems.
[2] C. Liang et al., “Generative AI-Driven Semantic Communica-
tion Networks: Architecture, Technologies and Applications,” Xi CHen is an adjunct professor at School of Computer Science,
arXiv preprint arXiv:2401.00124, 2023. McGill University, and also a lead AI Researcher at Proactive AI
[3] S. Javaid et al., “Leveraging Large Language Models for Integrated Lab, working on agentic AI. His experience and passion lie in a
Satellite-Aerial-Terrestrial Networks: Recent Advances and Future wide range of AI domains, including LLM agents, alignment, rea-
Directions,” arXiv preprint arXiv:2407.04581, 2024. soning, decision making foundation models, 5G/6G foundation
[4] P. Sahoo et al., “A Systematic Survey of Prompt Engineering models, AI for communications, Integrated Sensing and Commu-
in Large Language Models: Techniques and Applications,” nications (ISAC), autonomous driving, smart IoT, smart homes,
arXiv:2402.07927, 2024. smart systems, vehicle-to-everything, and so on. He achieved his
[5] H. Du et al., “Generative Al-Aided Joint Training-Free Secure PhD degree at School of Computer Science, McGill University. He
Semantic Communications via Multi-modal Prompts,” Proc. received both M.Eng. and B.S. degrees from Department of Elec-
ICASSP 2024-2024 IEEE Int’l. Conf. Acoustics, Speech and Sig- tronic Engineering, Shanghai Jiao Tong University.
nal Processing, 2024, pp. 12,896–900.
[6] W. X. Zhao et al., “A Survey of Large Language Models,” Hina Tabassum received the Ph.D. degree from the King Abdullah
arXiv:2303.18223, 2023. University of Science and Technology (KAUST). She is currently an
[7] Z. Lin et al., “Pushing Large Language Models to the 6g Associate Professor with the Lassonde School of Engineering, York
Edge: Vision, Challenges, and Opportunities,” arXiv preprint University, Canada, where she joined as an Assistant Professor, in
arXiv:2309.16739, 2023. 2018. She is appointed as a Visiting Faculty at University of Toronto
[8] M. Xu et al., “When Large Language Model Agents Meet 6g in 2024 and the York Research Chair of 5G/6G-enabled mobility
Networks: Perception, Grounding, and Alignment,” IEEE Wire- and sensing applications in 2023, for ve years. She is appointed
less Commun., 2024. as IEEE Comsoc Distinguished Lecturer for the term 2025-2026.
[9] Y. Liu et al., “Optimizing Mobile-Edge Ai-Generated Everything She is listed in the Stanford’s list of the World’s Top Two-Percent
(Aigx) Services by Prompt Engineering: Fundamental, Frame- Researchers in 2021-2024. She received the Lassonde Innovation
work, and Case Study,” IEEE Network, 2023. Early-Career Researcher Award in 2023 and the N2Women: Rising
[10] H. Zhou et al., “Large Language Model (LLM)-Enabled In-Con- Stars in Computer Networking and Communications in 2022.
text Learning for Wireless Network Optimization: A Case Study
of Power Control,” arXiv preprint arXiv:2408.00214, 2024. Xue Liu received PhD degree in computer science from the Uni-
[11] C. Hu et al., “Self-Rened Generative Foundation Models for Wire- versity of Illinois at Urbana–Champaign, USA. He is a William
less Trac Prediction,” arXiv preprint arXiv:2408.10390, 2024. Dawson Scholar (chair professor) and a full professor with the
[12] P. Esling and C. Agon, “Time-Series Data Mining,” ACM Com- School of Computer Science, McGill University, Montreal, Cana-
puting Surveys, vol. 45, no. 1, 2012, pp. 1–34. da. His research interests include computer and communication
[13] M. Chiang et al., “Power Control in Wireless Cellular Net- networks, real-time and embedded systems, cyber-physical sys-
works,” Foundations and Trends® in Networking, vol. 2, no. 4, tems and the IoT, green computing, and smart energy technolo-
2008, pp. 381–533. gies. He has published more than 200 research papers in major
[14] G. Barlacchi et al., “A Multi-Source Dataset of Urban Life peer-reviewed international journals and conference proceedings
in the City of Milan and the Province of Trentino,” Scientific in these areas and received several best paper awards.

106 IEEE Wireless Communications • August 2025