1

Generative AI-in-the-loop: Integrating LLMs and

GPTs into the Next Generation Networks
Han Zhang, Akram Bin Sediq, Ali Afana, and Melike Erol-Kantarci, Senior Member, IEEE

Abstract—In recent years, machine learning (ML) techniques on large datasets and designed to improve capabilities on
have created numerous opportunities for intelligent mobile net- specific tasks. Therefore, the performance of the model is
works and have accelerated the automation of network opera- highly dependent on the size and quality of the training dataset.
tions. However, complex network tasks may involve variables and
considerations even beyond the capacity of traditional ML algo- More importantly, under complex mobile communication envi-
arXiv:2406.04276v1 [cs.LG] 6 Jun 2024

rithms. On the other hand, large language models (LLMs) have ronments, the input variables and considerations of some tasks
recently emerged, demonstrating near-human-level performance are often beyond the capabilities of simple ML techniques [2].
in cognitive tasks across various fields. However, they remain A solution to this problem is to involve human interactions
prone to hallucinations and often lack common sense in basic in the ML-based automation process, a mechanism known
tasks. Therefore, they are regarded as assistive tools for humans.
In this work, we propose the concept of “generative AI-in-the- as ”Human in the Loop (HITL)” [3]. In this way, human
loop” and utilize the semantic understanding, context awareness, operators can utilize their semantic comprehension, problem-
and reasoning abilities of LLMs to assist humans in handling solving skills, and context awareness to deal with complicated
complex or unforeseen situations in mobile communication net- or unexpected situations. The downside of this mechanism is
works. We believe that combining LLMs and ML models allows that human intervention usually comes at a high cost and the
both to leverage their respective capabilities and achieve better
results than either model alone. To support this idea, we begin quality of human decisions highly depends on the level of
by analyzing the capabilities of LLMs and compare them with expertise and human reactions are slower relative to real-time
traditional ML algorithms. We then explore potential LLM-based decisions.
applications in line with the requirements of next-generation Large language models (LLMs), especially generalized pre-
networks. We further examine the integration of ML and LLMs, trained transformers (GPTs) have recently attracted significant
discussing how they can be used together in mobile networks.
Unlike existing studies, our research emphasizes the fusion of attention in the industry and academia. In addition to basic
LLMs with traditional ML-driven next-generation networks and language processing skills like semantic understanding or word
serves as a comprehensive refinement of existing surveys. Finally, generation, these models are also equipped with emerging
we provide a case study to enhance ML-based network intrusion abilities, such as instruction following and reasoning, as well
detection with synthesized data generated by LLMs. Our case as augmented abilities, such as interacting with others and self-
study further demonstrates the advantages of our proposed idea.
improvement. Therefore, LLMs have the potential to assist
Index Terms—Large language model, generalized pre-trained humans with tasks that require human intelligence [4].
transformer, 6G network, wireless communication, mobile net- Inspired by these thoughts, in this work, we propose the
works
concept of ”generative AI-in-the-loop” and utilize the abilities
I. I NTRODUCTION of LLMs to assist humans in handling complex or unforeseen
situations in mobile communication networks. Several existing
With the recent worldwide deployment of 5G and the studies have investigated the application of LLMs in commu-
ongoing standardization of 6G networks, mobile network nication systems. For instance, [2] discussed the integration of
evolution has been accelerating at a fast pace. In recent years, LLMs and GPTs within 6G systems and possible applications.
machine learning (ML) algorithms have emerged as effective [5] proposed an LLM-empowered multi-agent system and
and promising solutions in intelligent mobile networks thanks leveraged the interactions between agents to enhance the task-
to their superior ability to process data, provide insights solving capabilities in the 6G network. [6] discussed how
and take decisions. Techniques including classification, re- to combine LLMs with edge AI and meet users’ demands
gression, clustering, and reinforcement learning (RL) have with LLM-empowered code generation. These works usually
been adopted in application areas such as anomaly detection, make discussions at a generalized level, giving broad research
traffic prediction, handover control, and resource allocation. directions. Differently, our work focuses more on combining
The widespread use of ML techniques has accelerated the au- LLMs and ML models as solutions to specific problems in
tomation of selected tasks and laid the foundation for artificial mobile communication networks.
intelligence (AI)-native mobile communication networks [1]. The main contributions of this work are summarized as
The introduction of ML techniques also brought some follows:
challenges. Existing ML models are typically trained based
• We propose the concept of ”generative AI-in-the-loop”.
Han Zhang and Melike Erol-Kantarci are with the School of Electrical Engi- This concept involves leveraging LLMs to assist humans
neering and Computer Science, University of Ottawa, Ottawa, ON K1N 6N5, in managing and controlling complex network tasks that
Canada (e-mail: [email protected]; [email protected]).
Akram Bin Sediq and Ali Afana are with Ericsson, Ottawa, K2K 2V6, are typically beyond the capabilities of traditional ML
Canada(e-mail: [email protected]; [email protected]) techniques. This idea accelerates the full automation of
 2

Artificial intelligence (AI): Technologies that create machines capable of mimicking human intelligence.

Machine learning (ML): A subset of AI. Technologies that enable machines to learn from data and improve from experience without being
explicitly programmed

Traditional Statistical Large Language Models Artificial general

Neural networks (NN): A class (LLMs): Transformer-
Learning Methods intelligence (AGI):
of machine learning models based models on a large
inspired by the structure and A bigger vision
scale and pre-trained on that machines
functioning of the human brain. vast amounts of data.
Decision entail a broader
Random Trees GPT spectrum of
Generative AI (GAI): Evolved
Forests PaLM cognitive abilities
from ML. Algorithms that
LLAMA and go beyond
SVM contribute to various forms of
mimicking
content generation.
humans.

Fig. 1: Summary of concepts related to the topic of this work, including AI, ML, GAI, NN, LLM, GPT, and AGI.

network applications and takes a big step toward the ulti- amounts of data. With the popularity of LLMs, some compa-
mate goal of AI-native mobile communication networks. rable concepts have emerged to the public. The explanations
• Unlike existing studies, our research emphasizes the fu- for these concepts, including AI, ML, GAI, LLM, GPT, and
sion of LLMs with traditional ML-driven next-generation AGI, and their connections are summarized in Fig. 1. In this
networks. We argue that although LLMs and ML models work, we use the term ”traditional ML techniques” to refer
show great capabilities, they cannot replace the roles to the traditional statistical algorithms, reinforcement learning
of each other. Combining the generative and reasoning and others, used before the emergence of LLMs.
capabilities of LLM with the data analytics capabilities Benefiting from the remarkably large model size and en-
of ML can achieve better results than either model riched training data, LLMs possess a couple of critical capa-
alone can achieve. So we begin by analyzing the ca- bilities that have not appeared in previous NLMs. This makes
pabilities of LLMs and comparing them with traditional LLMs not only effective language processing tools but also all-
ML algorithms. We then explore potential integration by purpose task-solvers. The capabilities of LLMs are concluded
examining the life-cycle of the ML models. Our work also as follows [7]:
serves as a comprehensive refinement of existing surveys.
• In addition, a case study is given to enhance ML-based
• Understanding and generation: Language understanding
network intrusion detection with synthesized data gener- and generation are the two most basic abilities of LLMs.
ated by LLMs. It further demonstrates the advantages of It means that LLMs are capable of understanding human
our proposed idea. language without intentional instructions and generating
outputs in a similar form. Furthermore, recent tools are
The rest of the paper is organized as follows: Section II
combining LLMs with pre-trained multi-modal founda-
presents the concept of ”generative AI-in-the-loop” according
tion models. As a result, the understanding and generation
to LLMs’ capabilities and the requirements of the next-
capabilities of LLMs are not only limited to the text
generation networks. Section III discusses possible ways to
modality but also can be extended to other modalities
combine LLMs with traditional ML models and Section IV
like images, videos, and audio.
discusses the different deployments of LLMs and ML models.
• Reasoning: Reasoning abilities such as planning and log-
Finally, Section V presents a case study of a pre-trained LLM-
ical thinking enable LLMs to handle complex tasks like
enhanced network intrusion detection application and Section
solving mathematical problems and target decomposition.
VI concludes the paper.
These abilities can be promoted by skillful prompting
designs including chain-of-thought (CoT), tree-of-thought
II. G ENERATIVE AI- IN - THE - LOOP (ToT), self-consistency, and in-context learning (ICL).
In this section, we first explain the capabilities of LLMs and • Powerful knowledge base: The training process of LLMs
traditional ML models. Then, we discuss how to implement can be seen as a compression and abstraction of large
”generative AI-in-the-loop” and leverage LLMs and GPTs to amounts of training text. As a result, extensive knowledge
facilitate the automation of next-generation networks. is stored in the billions of parameters in LLMs and
serves as the basis for LLM-enabled problem-solving.
The knowledge base can also be continuously updated
A. Background and capabilities of LLMs and replenished with additional training data through fine-
LLMs refer to a group of transformer-based large-scale tuning.
statistical language models that are developed from neural • Social ability: Social ability means LLMs can commu-
language models (NLMs) and are pre-trained on massive nicate and interact with other LLMs or humans in an
 3

Benefits Concerns
• Ability to handle unforeseen problems: LLMs can
• Unstable results: The sampling methods make LLMs naturally
handle highly dynamic environments and deal with
randomized and have high stochastic outputs
unforeseen problems.
Pre-trained LLMs
• Less data dependency: Pre-trained LLMs can accomplish new
tasks with just a few contextual examples and do not • High Computation cost: The extremely large size of LLMs
need comprehensive training processes. result in a higher cost for training and fine-tuning, and limits
• Semantic understanding of input variables: LLMs the deployments on local servers.
can easily understand the meaning of different input
variables and their relations with the given task
without task-specific training.
• Explainable outputs and decisions: LLMs can
• Hallucinations: In some tasks, LLMs may generate outputs
generate explanations for outputs in human-like
that are nonsensical, inaccurate, or disconnected from inputs.
language forms, enhancing the explainability of
Data quality is critical to reducing hallucinations.
the decisions.
• Critical data demands: The quality of traditional ML models
• Smaller size and easier to train: Compared with LLMs,
greatly depends on the quality of training data. ML models are
traditional ML models usually have smaller sizes. That
Traditional vulnerable to inadequate training data, data quality issues, and
makes models easier to be trained and inferred locally.
ML models algorithmic biases.
• Deterministic output: Many of the existing ML models • Limited capabilities: Traditional ML models have restricted
are deterministic models and have higher reproducibility. capabilities and can be only used to process numerical inputs.
• Lack of flexibility: Traditional ML models are usually
• Data analytical ability: Traditional ML models are
trained for a single task. When the environment or task
more suitable for analyzing large datasets and
objectives change, the old ML model is no longer
capturing specific patterns in high-dimensional data.
applicable and needs to be retrained.

TABLE I: A comparison of the benefits and concerns using LLMs and traditional ML models

intelligible way. It is derived from a combination of un- of the next generation mobile networks. On the other hand,
derstanding and generation ability, reasoning ability, and the outputs of LLMs, especially GPTs, may be perceived
the powerful knowledge base of LLMs. This capability as demonstrating intelligence that is close to human levels.
facilitates the development of LLM-based AI agents for In a few existing studies [4] [8], LLMs have demonstrated
fully automated decision-making and control. comparable performance with that achieved by human experts
These above-mentioned capabilities serve as convincing ba- for some specific tasks, even though they can fail at simple
sis for the feasibility of ”generative AI-in-the-loop” in mobile tasks for humans or judgement that involves common sense.
networks. Nevertheless, in well-defined tasks, this promotes the possi-
bility of ”generative AI-in-the-loop”, for instance, LLMs to
assist humans for network management in slow timescales.
B. Traditional ML algorithms and LLMs
Although LLMs show strong capabilities, the deployments Fig. 2 shows an illustration of “generative AI-in-the-loop”
are usually limited by high cost, large scale, and unstable in the next-generation networks. As can be observed, there
results, which prevent LLMs from completely replacing tradi- are three optional levels of network configurations, human
tional ML models in many tasks. To make LLMs a good-fit level, LLM level, and ML level. All these three levels can
in ML-driven mobile networks, the benefits and concerns of observe desired information from the network environment
using LLMs and traditional ML models are first compared in and they can also interact with each other for collaborative
Table I. network management. Other than the traditional ML models,
As shown in this table, LLMs have the benefits of enhanced the ML level may also include optimization algorithms, rule-
capabilities, less data dependency, semantic understanding of based static models, or other small-size mathematical models.
input variables and explainable outputs and decisions. Tradi- This level directly collects data from the network environment
tional ML models have the benefits of smaller model size, in numeric format, conducts performance analysis, and selects
deterministic output and data analytical ability. This gives us desired information for decision-making. The parameters of
an insight into how they can be combined to solve mobile optimization algorithms and the rules of static models can be
network-related problems. set manually by human experts or tuned by the LLMs. The ML
models can be trained either online or offline with task-specific
data. The ML level holds the advantage of short inference time
C. Integrate LLMs into ML-driven next-generation network and low inference cost. As a result, applications with stricter
While ML techniques have greatly improved the efficiency timing requirements are suitable to be deployed at this level,
of network operations, LLMs come with additional capabilities including beam management, radio link monitoring, and user
that can further accelerate the development and full automation scheduling.
 4

Apply decisions
from three levels
Human level
Network environment Human
Human Expert supervision
intelligence Knowledge
Network
states Intent-based Explanations about
network control decisions
LLM level

Human
Knowledge base
supervision and
verification on
Multi-modal LLM/GPTs Memory local data
Multi-modal contextual data and
information contextual Decision making: Reasoning and planning
information
Network Model configurations and
Data Collection Function Design coordination
ML level
Performance analysis
Network
Task-specific Dataset
related data Select desired information

Data collection and processing

verification on
local data
Processed numerical Optimization Traditional Rule-based
network data algorithms ML models static models

Fig. 2: An illustration of “generative AI-in-the-loop” in the next-generation network. LLMs act as an intermediary between
human-level management and traditional ML and optimization algorithms in several ways: automate network control based on
semantic intentions, generate semantic-based explanations, and perform model and network management.

On top of the ML level, the LLM level is set. The LLM III. L EVERAGING LLM S FOR THE NEXT- GENERATION
level can acquire and comprehend multi-modal input from the NETWORKS
network environment. It also provides high-level guidance for In this section, we will first analyze some major issues in
the ML level by leveraging the strong knowledge base and ML-based mobile network management, and then introduce
memory within the context window for network task-related potential LLM-based applications according to the require-
planning, reasoning, and decision-making. For example, it can ments of next-generation networks.
guide the design of ML models or optimization algorithms,
decide the rules for rule-based static models, or perform A. Open issues in ML-based mobile network management
coordination between different models. In addition, consider- In this subsection, some open issues in ML-based mobile
ing the uneven development and instability of most LLMs, network management [9] and possible LLM-based solutions
supervisory mechanisms need to be established to verify for these issues are first discussed below:
the effectiveness of LLMs’ outputs and avoid hallucinations. 1) Scarcity of high-quality training data: ML-based network
Possible supervisory methods include human supervision or management approaches usually require large quantity of
verification of local data. high-quality data for training the models. However, public
The human level is set at the very top of the framework. datasets in this field, especially labeled datasets, are very
It is equipped with expert knowledge of mobile networks and scarce. Specifically, some datasets are outdated for network
can involve considerations related to human factors such as management because they were collected before some major
customer demands and commercial costs. The LLM level acts technological changes.
as an interface between the human level and lower levels. One possible solution is to use LLMs to augment or expand
It can translate intent-based network control into manageable datasets for model training. LLMs can also be leveraged to
tasks for ML models and generate explanations of ML layer identify anomaly samples in the dataset, improve the data
decisions for humans. quality, and synthesize new training data.
2) Limited Flexibility: Some traditional ML models are
Within the ”generative AI-in-the-loop” framework, the com- trained under the ideal network environment settings and
bination of LLMs and traditional ML models can capitalize they may not apply to realistic sub-optimal communication
on the benefits of both LLMs and traditional ML models and channels. Moreover, mobile networks are highly dynamic over
compensate for the weaknesses of the other. time considering the mobility of devices.
 5

LLMs can solve these problems by generating more in- specialized training. The agents are also expected to interact
telligent network management strategies. For instance, LLMs with each other and improve their performance on tasks
can extract constant latent representations from the dynamic through interactions with the environment and with other
network states and use them as new inputs for traditional ML agents in the absence of human intervention [6]. This can lead
models. to a agent-team-driven radio access network (RAN).
3) Security: Increasing diversity and complexity of mobile In the following sections, we will look more deeply into the
networks have brought new security concerns. Specifically, combination of ML and LLMs and discuss how they can be
deploying ML models opens mobile networks to external data simultaneously deployed in mobile networks.
and introduces new attacks such as data poisoning attacks
and membership inference attacks. This may exacerbate the IV. E NHANCE TRADITIONAL ML MODELS WITH LLM S
security challenges for the network.
In this section, we discuss the implementation of ”generative
To mitigate these security concerns, on the one hand, LLM-
AI-in-the-loop” from the perspective of traditional ML tech-
based intelligent security monitoring schemes can be used. On
niques. We illustrate how LLMs can be efficiently integrated
the other hand, some ML-based functions can be replaced by
into traditional ML model design.
LLMs since locally-run LLMs are better encapsulated and are
Fig. 3 shows seven different ways to enhance ML models
less vulnerable to traditional attack methods.
with LLMs. LLMs can use one of these ways or a combination
of all of these ways to play a supporting role to assist. In
B. Integrate LLMs into next-generation networks this figure, we divide the life cycle of traditional ML models
Inspired by the above analysis of issues and solutions, there into four stages: the requirement stage, the data processing
are three different ways to integrate LLMs into next-generation stage, the operation stage, and the model development stage.
networks: For different machine learning methods, such as supervised
1) Develop LLMs-based network functions: LLMs can be learning, unsupervised learning, and RL, the steps at the data
directly used to perform network functions that are usually processing stage can be different. In figure, we label the steps
developed with traditional ML models or rule-based static of different learning methods with different colors. LLMs
models. [10] verifies that LLMs can be used for precise label can play different roles at each stage, which is explained as
prediction after fine-tuning. [11] shows that LLMs are able follows.
to perform some simple decision-making tasks by alternating
reasoning and acting. These studies reflect the ability of LLMs A. Requirement stage
to carry out some network functions previously based on In the requirement stage, the model requirement analysis
traditional ML models. is performed to clarify the task requirements according to
In addition, LLMs can be integrated with ML models and the task description. This helps decide the number and type
used to enhance the performance of traditional ML models. of ML models needed for the given task. These designs are
In this way, the data scarcity and flexibility limitations of traditionally manually done with the help of human experts.
traditional ML models can be mitigated. Specific integra- However, with the emergence of LLMs, LLM-based model re-
tion methods are discussed in the following sections. Poten- quirement analysis can be performed to decompose semantic-
tial LLM-based network functions include traffic forecasting, based task descriptions into several small, manageable tasks to
anomaly detection, security monitoring, network automation, assist humans. LLMs can also select suitable models for each
and content compression. task. With the integration, LLMs take on a part of the model
2) LLM-assisted network application design: LLMs can also requirement analysis work that would have been manually
be used as effective tools while designing network applica- performed. We refer this as ”generative AI-in-the-loop” to
tions. For instance, they can be leveraged for code generation accelerate the automation of mobile networks.
and simulation system modeling. They can also be used to
decide ML model structures, hyper-parameters, or the rules
for developing static models. Additionally, LLMs can break B. Data processing stage
down large complex tasks into small manageable tasks, and In the data processing stage, two common steps are per-
distribute tasks to different ML models [2]. formed. In the first step, data is collected from mobile net-
3) Semantic understanding-empowered network manage- works and standardized. Next, data cleaning is conducted
ment: Considering the strong language understanding ability of to remove incomplete or anomalous samples from the data
LLMs, they can be used for semantic understanding in network set. LLMs can help with data cleaning by evaluating the
management. For instance, LLMs can extract configuration plausibility of the collected data samples.
information or optimization objectives from human-like input. In addition, there are two unique steps of supervised learn-
4) Build general-purpose AI agents for network nodes: ing models, feature selection and data labeling. In traditional
Another way to integrate LLMs into next-generation networks workflows, feature selection is usually performed by analyzing
is to build general-purpose AI agents on network nodes. By the correlation between features and the output. However, this
deploying LLMs on different network nodes, the agents are may lead to over-fitting problems and it is challenging to
expected to adapt to dynamic environmental changes and decide the statistical measures [12]. Instead, LLMs can be
perform different network tasks, even without intentional and used for data-free feature selection based on the semantic
 6

Data Processing Stage Integration of LLMs and

traditional ML models
Data Collection and Formatting Data Cleaning
Model Requirement 1. LLM-based model
Analyses Exploration and interaction with the environment requirement analysis
2. LLM-empowered data-
Requirement free feature selection
Feature Selection Data Labelling
Stage 3. LLM-based automatic
data labelling
4. LLM-based action space
LLM/GPTs-based reduction
model controller 5. LLM-based model
evaluation and hyper-
parameter optimization
Model Deployment Model Design Model Training 6. LLM-based ML model
design and code generation
Model Monitoring Model Optimization Model evaluation 7. LLM-based dynamic
Operation Stage Model Development Stage model status analysis and
lifecycle management
Common steps Unique steps of supervised learning Unique steps of reinforcement learning

Fig. 3: Different ways to enhance ML models with LLMs. The life cycle of ML models includes four stages: requirement
stage, data processing stage, operation stage, and model development stage. LLMs can be integrated into each stage.

understanding of given features. Data labeling is to add depending on different cases. Model size, inference time, and
informative labels to the raw data so that ML models can learn computation cost can be included as considerations for model
from it. The traditional data labeling approach is to ask humans management. LLMs can also be used to analyze the ML model
to recognize unlabeled data, which results in low scalability performance and evaluate if the model continues to function
and high expenses. In comparison, LLM-based automatic data as expected in realistic mobile network scenarios. If necessary,
annotation, or hybrid labeling with both human experts and it will initiate the life cycle of new ML models.
LLMs can automate the labeling process and lower the cost It is worth noting that non-ML models, such as optimization
[13]. algorithms and rule-based static models can also be included
Apart from this, RL models usually collect data by taking in such integration. For these models, the data processing stage
actions and receiving feedback from the environment. As a and the model development stage are not required, but they
result, an exploration and interaction with the environment can be developed on demand in the requirement stage and be
step is performed during this stage. In this step, LLMs can chosen by LLMs to perform tasks in the operation stage.
help with action space reduction and increase the efficiency of
the exploration. Furthermore, LLMs can also help synthesize V. D EPLOYMENTS OF ML MODELS AND LLM S IN MOBILE
training data when there is a shortage of available datasets. NETWORKS
In this section, we discuss how to simultaneously deploy
C. Model development stage traditional ML models and LLMs in mobile networks. As
The third stage is the model development stage. It includes shown in Fig. 4, there are three different deployment options:
model design, training, optimization, and evaluation. In the fully centralized deployment, mixed centralized and distributed
first two steps, the architecture of the ML model is decided deployment, and fully distributed deployment.
and the model is trained with the data prepared in the previous Fully centralized deployment means deploying both LLMs
stage. LLMs can be applied to model design through code and traditional ML models on the cloud side and combining
generation. After that, the ML model is evaluated and opti- them for network management. The benefit is that the powerful
mized. LLM-based model evaluation can help choose suitable computing and storage resources of the cloud server can
evaluation metrics, analyze the performance and automatically support the training and deployment of both LLMs and ML
tune the hyper-parameters. models. Moreover, centralized models have access to global in-
formation and can make more intelligent decisions. However,
the limitation is that the centralized structure lacks flexibility
D. Operation stage and scalability and is not suitable to handle a large number of
The last stage is the operation stage. In this stage, the ML local servers, for example, ultra-dense networks (UDNs). Also,
model is deployed at network nodes for model inference. It in this deployment, the cloud server will access sensitive data
also needs to be monitored in case of model decay, unexpected generated by base stations (BSs) and bring privacy concerns.
data and attacks. In this stage, LLMs can be used for dy- This deployment method is applicable when there is a small
namic model status analysis and life-cycle management. More number of BSs, and the controller wants to have centralized
specifically, LLMs decide which ML model should be used control of the whole network. It also fits the cases when the
 7

LLM/GPTs and traditional ML

LLM/GPTs on the cloud
models both on the cloud

1. Local model design

2. Model configuration
3. Model coordination

Instructions Interactions
between
LLM-based
agents

BS 1 BS 2 BS 1 BS 2

BS 1 BS 2 Each local server has a traditional ML model and a LLM Each local server has a traditional ML model.

Fully centralized deployment Fully distributed deployment Hybrid centralized and distributed deployment

Fig. 4: Three ways to combine LLMs with ML models. (a) Both LLMs and ML models are deployed at the center. They are
combined for network management. (b) Both LLMs and ML models are deployed in a distributed manner for Multi-Agent
interaction. (c) LLMs are deployed at the center for management, and ML models are deployed in a distributed manner for
local training and inference.

distributed BSs only have limited computation and storage and perform conflict management between different network
resources. Some possible applications include interference- applications.
aware radio resource management and power control, and
network slice provision. VI. A CASE STUDY
Fully distributed deployment means deploying both LLMs This section presents a case study on enhancing network
and traditional ML models on the local or edge site and intrusion detection using synthetic data generated by LLMs.
each local BS owns a LLM. The benefit is that the LLMs We consider the dataset proposed in [15] which is collected
and ML models are deployed closer to the data source. The from a cloud-based mobile communication system. In this
sensitive data can be kept locally to enhance data privacy. system, a malicious attacker can perform distributed denial-of-
In addition, fully distributed deployment has better scalability service (DDoS) attacks and generate malicious traffic. With the
since the workload is jointly undertaken by distributed BSs. labeled data, we train a convolutional neural network (CNN) to
The interactions between BSs are also promoted through perform network intrusion detection and decide the presence
multi-agent conversations [14]. However, the fully distributed of attacks in the network [15].
deployment has a much higher requirement on the resources Considering the scarcity of labeled data for the given task,
held by the local servers. The servers should be equipped we assume that there are only 20 available traffic data exam-
with powerful computing and storage resources to support both ples. Among these examples, 10 examples are collected from
LLMs and ML models. Additionally, the distributed structure benign network traffic and 10 examples are collected from
will introduce parallelization and synchronization steps, and malicious network traffic. To improve the intrusion detection
expose models to more vulnerabilities. accuracy, additional traffic data examples are synthesized by
In conclusion, this deployment method applies to scenarios GPT-3.5 and used for CNN training. The prompt for network
with powerful local edge servers and where interactions are traffic generation consists of four parts: task description, ex-
needed between BSs. Some suitable applications include traffic amples listing, data explanation, and output formatting. In the
prediction and user association. first part, the task for GPT-3.5 is described as ”generating some
Hybrid centralized and distributed deployment combines the data to train an ML model for network intrusion detection”.
previous two deployment methods. In this case, LLMs are de- Next, available training data is listed as examples for data
ployed on the cloud while traditional ML models are deployed generation. In the third part, an explanation of the meaning
on local servers. This method incorporates the benefits of both of each value in the dataset is given to GPT-3.5 and the
fully centralized deployment and fully distributed deployment. semantic information can assist in data generation. Finally,
The cloud server is usually equipped with powerful resources output formatting is given to limit the amount of generated
to support the needs of LLMs. The distributed ML models data and make the generated data more manageable.
are supported by local devices and are customized for task- In our simulations, we first explore the impact of the amount
oriented applications. As a result, resources can be rationally of synthetic data on the performance. Fig. 5a and 5b show the
exploited and data privacy can be protected. The only concern performance of TCP ACK attack detection while using GPT-
is that the communication between local ML models and 3.5 to generate different numbers of synthetic examples. We
the cloud-based LLMs will result in extra communication compare the accuracy and the F1-score while only using the
overhead. real data, only using the synthetic data, and using mixed data
With this deployment, LLMs can be used for distributed ML of both synthetic data and real data. As it can be observed, the
model management. They can analyze user requirements, plan synthetic data can help improve the detection accuracy from
tasks, and leverage ML models for execution. For example, 71.3% to above 80%. Meanwhile, an increase in the number of
LLMs can select the appropriate scheduling algorithm models synthesized examples does not necessarily improve detection
 8

90% 0.9

85% 0.85

80% 0.8
Accuracy increased
Accuracy

Accuracy increased

F1-Score
75% with synthetic data 0.75 with synthetic data
70% 0.7
Synthetic data only Synthetic data only
65% Mixed data 0.65 Mixed data
Real data only Real data only
60% 0.6
20 40 80 160 20 40 80 160
Number of synthetic examples Number of synthetic examples
(a) Accuracy of network intrusion detection under different numbers of synthetic (b) F1-score of network intrusion detection under different numbers of synthetic
examples. examples.
Before self-evolution After self-evolution
0.9 Synthetic data only Mixed data Real data only
100%
0.7
Values

90%

Accuracy
0.5
80%
0.3
70%
0.1
Accuracy - F1 - Synthetic Accuracy - F1 - Mixed data 60%
Synthetic data data only Mixed data TCP ACK TCP SYN TCP push UDP flood TCP FIN
only attack attack attack attack attack
Data source Attack type

(c) Accuracy and F1-score of network intrusion detection using the synthetic data (d) Accuracy of network intrusion detection using synthetic data generated by
before self-evolution and after self-evolution. GPT-3.5 under different types of attacks.

Synthetic data only Mixed data Real data only

1

0.9
F1-Score

0.8

0.7

0.6
TCP ACK TCP SYN TCP push UDP flood TCP FIN
attack attack attack attack attack
Attack type

(e) F1-score of network intrusion detection using synthetic data generated by GPT-
3.5 under different types of attacks.

Fig. 5: Accuracy and F1-score of network intrusion detection using synthetic data.

performance. This reveals that the generation ability of LLMs confusing data. Considering the current limited capabilities of
is limited and the quality of synthetic data may decrease as GPT-3.5, the quality of synthesized data should be verified by
more samples are generated. the real dataset before being used for CNN training.
On the other hand, we notice that synthetic data generated
by GPT-3.5 sometimes has instability issues. To further ensure Finally, Fig. 5d and Fig. 5e show the accuracy and F1-
high-quality data generation, a self-evolution can be added by score of network intrusion detection using synthetic data
using prompts like ”These examples are not accurate enough generated by GPT-3.5 under different types of attacks while 80
to train ML models. Can you generate better data”. synthetic examples are generated. For the TCP ACK attack,
Fig. 5c shows a comparison of the synthetic data before the accuracy and F1-Score are improved by 22.1%, and for
and after self-evolution for the TCP FIN attack. As can be the TCP FIN attack, the accuracy and F1-score are improved
observed, before self-evolution, the CNN model trained with by 28.7%. In addition, synthetic data enhances the detection
synthetic data shows low accuracy and F1-score. It implies of different attacks differently. This is because the quality
that the quality of the generated data is far inferior to the of the real data, the difficulty of the task, and the ability
original data. In contrast, after self-evolution, the data quality of LLMs to recognize attack patterns in the given examples
is improved. It is worth noting that self-evolution does not vary for different attacks. In conclusion, the simulation results
always promote performance. When GPT-3.5 has already demonstrate that synthesizing data with LLMs for ML model
output high-quality synthetic data, it may repeat the previously training is an effective way to solve data scarcity issues and
generated data. In a few cases, it also outputs invalid and enhance model performance for mobile communication tasks.
 9

VII. C ONCLUSION [13] S. Wang, Y. Liu, Y. Xu, C. Zhu, and M. Zeng, “Want
to reduce labeling cost? GPT-3 can help,” arXiv preprint
ML techniques offer numerous opportunities for intelligent arXiv:2108.13487, 2021.
mobile networks and accelerated automation. On the other [14] Q. Wu, G. Bansal, J. Zhang, Y. Wu, S. Zhang, E. Zhu, et al.,
hand, LLMs and GPTs have recently garnered significant “Autogen: Enabling next-gen llm applications via multi-agent
attention due to their outstanding performance in cognitive conversation framework,” arXiv preprint arXiv:2308.08155,
tasks across various fields. Inspired by these thoughts, we 2023.
[15] B. Farzaneh, N. Shahriar, A. H. Al Muktadir, and M. S.
propose that combining the strengths of both LLMs and Towhid, “DTL-IDS: Deep transfer learning-based intrusion
ML models can yield a synergistic effect, making the whole detection system in 5G networks,” in 2023 19th International
greater than the sum of its parts. In this work, we discuss Conference on Network and Service Management (CNSM),
the strengths of ML models and LLMs in different tasks and IEEE, 2023, pp. 1–5.
explore how to effectively combine them to address mobile
network-related challenges. We also provide a case study to
demonstrate the advantages of our proposed idea which utilizes
generation abilities of LLMs to improve the accuracy of ML-
based intrusion detection. In the future, we plan to integrate
LLMs into more complex and advanced ML-driven network
applications.

ACKNOWLEDGMENT
This work has been supported by MITACS and Ericsson
Canada, and NSERC Canada Research Chairs program.

R EFERENCES
[1] M. Iovene, L. Jonsson, D. Roeland, M. D’Angelo, G. Hall,
M. Erol-Kantarci, et al., “Defining AI native: A key enabler
for advanced intelligent telecom networks,” Ericsson, Tech.
Rep., 2023.
[2] S. Tarkoma, R. Morabito, and J. Sauvola, “AI-native intercon-
nect framework for integration of large language model tech-
nologies in 6G systems,” arXiv preprint arXiv:2311.05842,
2023.
[3] Q. Yan, W. Chen, and H. V. Poor, “Big data driven wireless
communications: A human-in-the-loop pushing technique for
5G systems,” IEEE Wireless Communications, vol. 25, no. 1,
pp. 64–69, 2018. DOI: 10.1109/MWC.2018.1700209.
[4] C.-H. Chiang and H.-y. Lee, “Can large language mod-
els be an alternative to human evaluations?” arXiv preprint
arXiv:2305.01937, 2023.
[5] F. Jiang, L. Dong, Y. Peng, K. Wang, K. Yang, C. Pan, et al.,
“Large language model enhanced multi-agent systems for 6G
communications,” arXiv preprint arXiv:2312.07850, 2023.
[6] Y. Shen, J. Shao, X. Zhang, Z. Lin, H. Pan, D. Li, et al.,
“Large language models empowered autonomous edge AI
for connected intelligence,” IEEE Communications Magazine,
2024.
[7] S. Minaee, T. Mikolov, N. Nikzad, M. Chenaghlu, R. Socher,
X. Amatriain, et al., “Large language models: A survey,” arXiv
preprint arXiv:2402.06196, 2024.
[8] G. Suri, L. R. Slater, A. Ziaee, and M. Nguyen, “Do large
language models show decision heuristics similar to humans?
a case study using GPT-3.5.,” Journal of Experimental Psy-
chology: General, 2024.
[9] Y. Liu, H. Du, D. Niyato, J. Kang, Z. Xiong, D. I. Kim, et
al., “Deep generative model and its applications in efficient
wireless network management: A tutorial and case study,”
arXiv preprint arXiv:2303.17114, 2023.
[10] Z. Li, X. Li, Y. Liu, H. Xie, J. Li, F.-l. Wang, et al., “Label su-
pervised llama finetuning,” arXiv preprint arXiv:2310.01208,
2023.
[11] S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan,
et al., “React: Synergizing reasoning and acting in language
models,” arXiv preprint arXiv:2210.03629, 2022.
[12] J. Brownlee, “How to choose a feature selection method for
machine learning,” Machine Learning Mastery, vol. 10, 2019.