Zhao et al. J. Wirel. Commun. Netw.

(2026) 2026:39 EURASIP Journal on Wireless

[Link]
Communications and Networking

RESEARCH Open Access

Scheduling and resource allocation in 5G

SCMA cellular networks using AI‑based methods
Mincheng Zhao1* , Mingqing Han2, Jiyuan Pan1, Ligang Du1, Yongfeng Yan1 and Fuqiang Li1*

*Correspondence:
mczhao1986@[Link]; fqli@163.
Abstract
com The deployment of 5G networks has incorporated advanced multiple access technolo-
1
The 20th Research Institute gies like sparse code multiple access (SCMA) to address growing demands for high-
of CETC, Xi’an, Shaan xi, China speed connectivity and massive device access. As a Non-Orthogonal Multiple Access
2
China Mobile Communications
Group Shaanxi Co. Ltd, Xi’an, technique, SCMA enables multiple users to share identical time-frequency resources
Shaan xi, China through sparse codebook-based multiplexing. Nevertheless, achieving efficient
scheduling in SCMA networks remains challenging due to the inherent complexi-
ties in dynamic resource allocation. This paper proposed two artificial intelligence-
based approaches for resource scheduling in 5G SCMA networks: a multi-agent
deep reinforcement learning (MARL)-based approach and a large language model
(LLM)-empowered methodology. We systematically investigate these AI techniques
to develop adaptive resource scheduling policies capable of responding to diverse net-
work conditions. Simulation results validate that the proposed MARL-based and LLM-
based schedulers not only effectively learn optimal scheduling strategies but also out-
perform conventional algorithms, particularly in terms of system throughput and user
fairness metrics.
Keywords: Scheduling and resource allocation, Sparse code multiple access, Multi-
agent deep reinforcement learning, Large language model, Artificial intelligence

1 Introduction
The fifth generation (5G) of cellular networks is designed to support a wide range of
applications, including enhanced mobile broadband (eMBB), ultra-reliable low-latency
communications (URLLC), and massive machine-type communications (mMTC).
SCMA [1] is a promising multiple access technique for 5G networks, enabling high spec-
tral efficiency and massive connectivity by enabling multiple users to share the same
time-frequency resources through sparse codebooks. However, despite its advantages,
SCMA presents significant challenges in resource scheduling and interference manage-
ment due to its non-orthogonal nature.
Traditional scheduling algorithms, such as those based on orthogonal frequency-
division multiple access (OFDMA), are not well-suited for SCMA due to its unique
characteristics. The network operates in a time-slotted manner, with each time slot
corresponding to a scheduling decision. The objective of the SCMA scheduler [2] is to

© The Author(s) 2026. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0
International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you
modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of
it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise
in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted
by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy
of this licence, visit [Link]
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 2 of 26

allocate codebooks and channel resources to user equipment (UEs) in a manner that
maximizes network performance metrics, including throughput and user fairness.
In wireless networks, the wireless channels are in a time-varying state, and tradi-
tional static algorithmic mechanisms are inadequate for addressing the complexities
of such environments. Consequently, researchers [3, 4] use deep learning methods to
tackle the challenges associated with wireless scheduling algorithms in cellular net-
works. Deep reinforcement learning (DRL) is proven to be a powerful tool for solv-
ing complex resource allocation optimization problems in dynamic environments.
For the complex multi-resource allocation problem, the authors [5] proposed uses
multi-agent deep reinforcement learning (MARL) method in D2D networks, where
each D2D pair acts as an independent agent making local decisions on subchannel
and power allocation.
Furthermore, with the rapid development of LLMs such as ChatGPT [6] and Deep-
Seek [7], researchers have begun exploring their innovative applications in wireless
communication resource allocation. By leveraging their powerful capabilities in natu-
ral language processing, pattern recognition, and decision optimization, LLMs offer a
new research paradigm for traditional resource allocation algorithms [8].
In this paper, we explore the adoption of multi-agent deep learning and LLM meth-
ods to address the multiple resource allocation challenges associated with wireless
scheduling algorithms in SCMA cellular networks. We propose a MARL-based and
LLM-empowered scheduling framework for 5G SCMA networks, aiming to optimize
key performance metrics such as throughput and user fairness. Our main contribu-
tions are highlighted as follows.

• Downlink SCMA scheduling in 5G cellular network is formulated as the multi-

objective optimization problem under the condition of link adaptation.
• A novel MARL method and a LLM-based methods are proposed to address the
issue of the SCMA resource scheduling. Firstly, we assign appropriate functional
representations to the state, action, reward, and use optimal MARL method to
resolve the integrated SCMA resource allocation problem. Secondly, for the LLM-
based method, we explore various prompting strategies to optimize resource
scheduling performance.
• By evaluating and comparing the traditional methods, we could demonstrate the
MARL and LLM learning capability and how expert knowledge to help improve
performance.

Some preliminary results of our proposed algorithms are reported in [9]. This paper
provides more comprehensive MARL algorithm, LLM prompts design and experimen-
tal results. Firstly, a more detailed description of MARL algorithm is provided, includ-
ing the L-W greedy algorithm. Secondly, we provide a more comprehensive description
of LLM-based scheduling algorithm, along with the design and a comparative analysis
of two different prompts. Thirdly, We conducted a more comprehensive experimental
validation, which included add the reference comparative algorithm MASR_SCMA, per-
forming a comparative analysis of the scheduling effectiveness of different LLMs with
different prompts, and analyzing system capacity, among other aspects.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 3 of 26

The rest of this paper is organized as follows. Section 2 introduces the related work.
Section 3 introduces the proposed framework, the proposed MARL method and
LLM-empowered method. Simulation setup and results are presented in Sects. 4 and
5. We also discuss the challenges and open questions in Sect. 6. Followed by conclud-
ing remarks in Sect. 7.

2 Literature review
In resource allocation for wireless communication systems, parameters such as the
channel and transmit power are adjusted to optimize network performance, includ-
ing maximizing system throughput and fairness while satisfying various constraints.
Finding the optimal strategy is typically formulated as an optimization problem,
which can be addressed using conventional optimization-based approaches, DRL-
based approaches and LLM-based approaches as follows (Fig. 1):

• Optimization-based approach: The formulated optimization problem is tackled

analytically. In general, the optimization problems is complex and hard to solve,
often involving integer-valued control parameters and non-convex functions. Con-
sequently, various mathematical approaches are employed to transform and simplify
the problem, making it more tractable. The optimal solution is typically obtained
through priority-based ranking algorithms or iterative methods.
• DRL-based approach: The optimal resource allocation is approximated using the
output of a specially designed DRL, which generates the resource allocation strategy
based on the current conditions of the wireless network. To achieve this, the DRL
must be trained, which can be achieved either through supervised learning that uti-
lizes optimal resource allocation as labeled data or through unsupervised learning
that does not require labeled data. Although training the DRL can be time-consum-
ing, inference from the trained DRL can typically be executed in milliseconds level
latency as studied in [10].

Fig. 1 The comparison between Optimization-based, DRL-based and LLM-based methods for wireless
resource allocation problem
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 4 of 26

• LLM-based approach: LLM-based resource allocation presents a promising solution

to the limitations of conventional schemes. In this approach, resource allocation can
be determined by the LLM using a few-shot learning method, which eliminates the
need for retraining the LLM for specific scenarios. The LLM acts as a general opti-
mizer, capable of addressing a wide range of resource allocation problems. The ben-
efits of using LLMs in resource allocation can be summarized as follows: First, LLM-
based schemes can effectively handle various multimodal data. By leveraging such
multimodal data in wireless communication systems, more efficient and proactive
radio resource allocation can be achieved. Second, unlike DRL-based schemes that
require extensive training, LLMs can operate without task-specific training. Conse-
quently, LLMs can easily adapt to a wide range of objectives and topologies.

2.1 Optimization‑based wireless resource scheduling

In modern wireless communication systems, efficient resource allocation is critical to
achieving high performance and ensuring fairness among users. As the demand for
mobile data continues to grow exponentially, the need for sophisticated scheduling algo-
rithms that can balance throughput maximization and fairness has become increasingly
important. Among the various scheduling strategies proposed in the literature, the pro-
portional fair (PF) scheduling algorithm [11] has emerged as a widely adopted solution
due to its effectiveness in balancing these two competing objectives. The PF schedul-
ing algorithm was originally designed to optimize resource allocation in code-division
multiple access (CDMA) systems, such as 3 G networks. Its fundamental principle is
to allocate resources to users based on their instantaneous channel conditions in rela-
tion to their average throughput. By prioritizing users with favorable channel conditions
while ensuring that all users receive a fair share of resources over time, the PF algorithm
achieves a balance between maximizing system throughput and maintaining fairness.
Despite its widespread adoption, the PF scheduling algorithm encounters several chal-
lenges in modern wireless networks. In 5G and beyond-5G (B5G) systems, the increas-
ing complexity of network architectures, the heterogeneity of user requirements, and the
necessity to support diverse services (e.g., URLLC, eMBB, and mMTC) have necessi-
tated further enhancements to the traditional PF algorithm. Additionally, the integration
of multi-antenna technologies, such as massive multiple-input multiple-output(MIMO)
and the deployment of ultra-dense networks (UDNs), has introduced new dimensions to
the resource allocation problem, requiring more advanced scheduling strategies.
SCMA [1] is a promising multiple access technique for 5G networks, enabling high
spectral efficiency and massive connectivity by enabling multiple users to share the same
time-frequency resources through sparse codebooks. PF algorithm which works based
on the CDMA or OFDMA network, are not well-suited for SCMA due to its unique
characteristics. The researchers begin to tackle the challenges associated with wire-
less resource allocation in 5G SCMA networks, mainly focus on improving the system
throughput, fairness and energy efficiency (EE). The objective of the SCMA scheduler [2]
is to allocate codebooks and channel resources to UEs using matching game method in
a manner that maximizes network performance metrics, including throughput and user
fairness. A greedy and Lagrange dual optimal method [12] is proposed to allocate the RB
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 5 of 26

and power resource in order to maximize SCMA system throughput. The impact of the
factor graph matrix on the average sum rate was studied in [13] by assuming the known
link distances between users and the base station (BS), and a low-complexity iterative
algorithm is proposed to design the optimal graph matrix which maximizes the average
sum rate of the SCMA systems for the general parameters. The authors [14] studied to
maximize the NOMA system’s sum rate and energy efficiency (EE) using the proposed
iterative water-filling solution. An iterative algorithm in [15] was utilized in the resource
assignment problem to maximize energy efficiency.

2.2 DRL‑based wireless resource scheduling

The rapid advancement of DRL has revolutionized the field of artificial intelligence, ena-
bling agents to learn complex behaviors through interactions with their environments.
From mastering board games like Go and chess to controlling autonomous vehicles,
DRL has demonstrated remarkable success in single-agent settings. However, many real-
world applications involve multiple agents operating in shared environments, where
their actions are interdependent and the system’s dynamics are influenced by the col-
lective behavior of all agents. This has spurred significant interest in MARL, a subfield
that extends DRL to scenarios involving multiple interacting agents. In SCMA systems,
a single resource block (RB) can carry superimposed signals from multiple users. The
allocation among multiple resources involves mutual constraints and potential impacts
(including allocation limitations and signal interference), MARL methods are required
to address this more complex resource allocation.
In the DRL algorithm for wireless resource allocation [3, 4], the state of the network
at each time slot is represented by a high-dimensional vector that includes information
such as channel state information (CSI), buffer status, and historical performance met-
rics. The action space consists of all possible codebook and power allocations to UEs.
The reward function is designed to reflect the desired performance objectives, such as
maximizing throughput while ensuring fairness among UEs. The authors [16] use deep
reinforcement learning to make appropriate radio resource allocation decisions in pro-
active vehicular networks with ultra-low latency. A simple adaptive MAC strategy [17]
based on the Q-learning has been proposed to reduce the number of collisions, thereby
decreasing packet latency in CSMA networks. A DRL-based method is proposed [18]
to maximize the eMBB throughput while satisfying the latency requirements of the
URLLC slice. The authors [19] proposed a deep Q-learning-based SCMA, in which
codebooks are adaptively constructed to minimize bit error rate (BER), maximizing
throughput within the constraints of reliability and latency. In order to resolve the radio
resource management problem between intra-subnetwork and inter-subnetwork due to
the dynamic mobility of wireless subnetworks, the authors [20] propose a novel effec-
tive intelligent radio resource management method using MARL, which only needs the
sum of received power, named received signal strength indicator (RSSI), on each channel
instead of channel gains. Moreover, researchers also adopt DRL series methods for wire-
less network slicing management and transmission service optimization. The authors
[21] comprehensively reviews AI-driven methods, particularly machine learning (ML)
and deep reinforcement learning (DRL), applied to 5G network optimization, focus-
ing on resource allocation, traffic management, and network slicing. The authors [22]
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 6 of 26

present a consolidated review of the state of the art in network slicing resource manage-
ment modules and network slicing-enabled key industrial use cases in 5G core network
using AI-based methods, such as RL method. The author [23] proposes a DRL-based
rate adaptation algorithm for adaptive 360-degree video streaming, which maximizes
the quality of experience for viewers by adjusting the transmitted video quality to the
time-varying network conditions.

2.3 LLM‑based wireless resource scheduling

LLMs represent a class of artificial intelligence systems based on deep neural networks,
predominantly utilizing the transformer architecture [24] to achieve state-of-the-art per-
formance in natural language understanding and generation. These models are trained
through self-supervised learning on extensive textual corpora, enabling them to acquire
broad and general reasoning capabilities—specifically, increasing model parameter
size, dataset breadth, and computational resources leads to emergent abilities, includ-
ing in-context learning, mathematical-logical analysis, and multi-step problem-solving.
By leveraging its complex reasoning and analytical capabilities, it can obtain resource
allocation feedback at near-instant levels, making it suitable for coarse-grained resource
scheduling scenarios. However, LLM deployment is cost-intensive.
Among the notable innovations in efficient LLM architectures, DeepSeek demon-
strates significant advancements through its compute-aware design and dynamic infer-
ence optimization. The model [25] employs hierarchical sparse attention with adaptive
computation paths, reducing floating point operations per second (FLOPs) by 35%
compared to dense Transformers while maintaining 98% of performance on language
understanding tasks. R1’s architecture features token-level early exiting and mixture-of-
experts [26] components.
Many researchers have begun to study the LLM-empowered wireless network opti-
mization. The authors [8] propose WirelessLLM, a comprehensive framework for
adapting and enhancing LLMs to address the unique challenges and requirements of
wireless communication networks. The authors also investigate the enabling technolo-
gies to build WirelessLLM, including prompt engineering, retrieval augmented genera-
tion, multimodal pre-training, and domain-specific fine-tuning. The LLM-based power
scheduler [27] is proposed to minimize BS transmission power while simultaneously sat-
isfying average data rate constraints. The authors [28] present an illustrative example of
a simple resource allocation problem that aims to maximize either spectral efficiency
(SE) or EE, and design an LLM-based resource allocation framework. The authors pro-
pose a large language model [29] for resource allocation optimization (LLM-RAO), a
novel approach that leverages the capabilities of LLMs to address the complex resource
allocation problem while adhering to quality of service (QoS) constraints. The article
[30] introduces the LLM-empowered hierarchical RIC (LLM-hRIC) framework, which
enhances real-time collaboration between RICs in O-RAN. To address the problem of
non-convex resource allocation in wireless communication systems, the authors propose
LLM-OptiRA [31], the first framework that LLMs to automatically detect and transform
non-convex components into solvable forms, thereby enabling fully automated resolu-
tion of non-convex resource allocation problems in wireless communication systems.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 7 of 26

3 Methods and methodology

3.1 System description
We consider a 5G SCMA network with a single BS serving multiple UEs. The network
operates in a time-slotted manner, with each time slot corresponding to an individual
scheduling decision. The objective of the scheduler is to allocate codebooks and chan-
nel resources to UEs in a manner that maximizes network performance metrics such as
throughput and user fairness.
In this paper, we assume that each SCMA layer represents a user, and the base sta-
tion can allocate a maximum of J codebooks to each user. Each codebook contains L
codewords of length K, with each codeword consisting of a certain number of non-zero
elements dV . The bit-level information transmitted to the users can be mapped to the
corresponding sparse codewords in the codebooks. The base station has access to the
CSI and allocates wireless resources based on both the assigned codebooks and the cur-
rent channel conditions of the users.
Figure 2 illustrates the proposed single cell cellular system model. The BS delivers the
requested application servers to the users via the backbone network and the wireless cel-
lular network. The BS is equipped with an AI-based scheduler powered by either MARL
or an LLM. This AI-enabled scheduler doesn’t break into the established conventional
modules, only focuses specifically on optimizing wireless resource allocation. The sys-
tem includes multiple resource blocks (RBs). The BS collects channel state information
and feedbacks to the MARL or LLM scheduling module, which installed to cooperate
with the scheduling module to achieve the optimum.

Fig. 2 The overall architecture for the proposed AI-based methods

Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 8 of 26

3.2 Problem formulation
We consider a centrally controlled cellular downlink communication system, which is
used in 5G cellular Network. A set of users m={1,…,M} requests services and the BS trans-
mits messages to these users via both the backbone network and the wireless cellular net-
work. The system operates over a total available bandwidth B, which is equally divided into
N={1,…,N} resource block groups, each with 12 consecutive subcarriers. All the chan-
nels are assumed to exhibit block fading characteristics. For each user m ∈ M and each
subchannel n ∈ N , the transmit power and the channel gain for user k in subchannel n are
denoted as pm,n and hm,n, respectively, and σ 2 represents the ambient noise variance. The
communication system is assumed to employ an M-QAM modulation scheme.
The SCMA resource allocation problem can be mathematically modeled as:
M

Problem 1 : maximize Rm + γ F , (1)
m=1

s.t. Eq. (10) − (11), (2)

where Rm indicates the average rate of user m, F represents the fairness (indicated
by Jain’s fairness, which is popular in wireless scheduling index as [3, 4]) index. The
throughput in the reward function is the received instantaneous throughput, hence the
UE with the best channel condition tends to be chosen by considering throughput only.
The fairness component in the reward focuses on equity in average throughput among
different users. Consequently, resources may be allocated to users whose channel condi-
tions are not optimal, in order to ensure fairness, γ indicates the weight parameter which
controls the trade-off between rate and fairness. Jain’s fairness F can be calculated as:
M 2
[ m=1 Rm ]
F= M 2 (3)
M m=1 [Rm ]

Considering that γ is a number with a value range of [0, 1], we divided the rate Rm by 1
million when calculating resource allocation efficiency to bring its value range close to 1,
thereby effectively balancing the rate value with fairness and obtaining a more reason-
able reward value.
According to the resource allocation principle [32] of SCMA, a subcarrier can be allo-
cated to a maximum of dJ users, and a user can simultaneously receive data from a maxi-
mum of dV subcarriers. The total rate of user m across all subcarriers can be expressed as:
N

Rm = µm,n rm,n
n=1
N
(4)
 B  βpm,n |hm,n |2 
= µm,n log2 1 +
N σ 2 + Im,n
n=1

µm,n ∈ {0, 1} (5)

where µm,n = 1 indicates that RB n is allocated to user m and µm,n = 0 indicates that
RB n is not allocated to user m. rm,n represents the transmission rate of user m on RB n.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 9 of 26

Assuming the base station employs M-QAM modulation, Im,n represents the interfer-
ence received by user m on subcarrier RB n from other users. User m’s feasible rate on
subchannel n can be expressed as:

B  βpm,n |hm,n |2 
rm,n = log2 1 + . (6)
N σ 2 + Im,n

where β is a constant related to a targeted BER through the relation β = ln(5∗BER)

−1.5
[33].
Furthermore, users with better channel conditions can adopt higher-order modulation
schemes to achieve higher data rates. Assuming the base station allocates equal transmit
power to different SCMA layers, the received power from other users’ data is the same
for user m. And Im,n represents the interference received by user m, Im,n can be calcu-
lated as:

Im,n = pi,n |hi,n |2 (7)
i∈{Sn |(|hi,n |2 <|hm,n |2 )}

which is user m receives from other users Sn |(|hi,n |2 < |hm,n |2 ) on resource block n
[32]. Since RB n can be utilized by a subset of users, the signal of any user i will cause
the interference to other users in RB n. To demodulate the target signal, usually SCMA
receivers utilize the successive interference cancellation (SIC) decoding [34]. It firstly
decodes the signal of the user with better channel conditions, subtracts it, and then
decodes its own signal. Therefore, when user m decodes its own signal, the interference
it experiences comes from all users with poorer channel conditions than itself. In the
base station, the data transmitted to each user are mapped to the corresponding sparse
code of the codebook. The data from all m SCMA layers are multiplexed onto N shared
resource blocks. The signal on the resource block can be expressed as:
M

yn = hm,n xm,n + wn (8)
m=1

where hm,n represents the channel coefficient of user m on resource block n, xm,n denotes
the codeword information of user m on resource block n, and wn represents the Gauss-
ian white noise on resource block n.
In SCMA resource allocation, a subcarrier can be allocated to a maximum of dJ users,
which is expressed as:
M

µm,n ≤ dJ , ∀n ∈ N (9)
m=1

Additionally, a user can simultaneously receive data from a maximum of dV subcarriers,

satisfying:
N

µm,n ≤ dV , ∀m ∈ M (10)
n=1

According to [1, 35], dV is the number

 of RBseach user occupies, which is set to dV . The
K −1
best dJ can be calculated as dJ = , where K is usually set to the number of
dV − 1
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 10 of 26

 
N
orthogonal RBs N. Moreover, the best number of users is calculated as M ∗ = .
dV
The total transmit power of the base station must not exceed the rated value P, so the
following constraint must also be satisfied:
M 
 N
pm,n ≤ P
m=1 n=1 (11)
pm,n ≥ 0, ∀m ∈ M, n ∈ N

Our goal is to determine the optimal binary variables µm,n for each episode. Moreover,
Eq. (1) is a complex non-convex problem to stably optimize a fixed objective for system
throughput or fairness, due to the binary constraint as well as the existence of the inter-
ference term, which is inherently challenging to solve as discussed in [32]. Furthermore,
the static policy can not adapt to the complex network conditions, sometimes it is even
opposed to the objective. Figure 3 illustrates the detailed framework.
The problem is NP-hard, which could be proved as follows.

Definition 1 The Multiple Knapsack Problem (MKP): Given a knapsack with capac-
ity V and N types of items, where each type n ∈ 1, . . . , N has a maximum availability of
sn items. Each item m of type n possesses a value wm,n and a volume vm,n. The problem
requires selecting quantities of items (up to sn for each type n) to pack into the knapsack

Fig. 3 RB allocated by different AI agents for SCMA. Wireless resource is allocated in each episode, and each
AI agent is responsible for the allocation of one RB
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 11 of 26

such that: (1) the total volume does not exceed V, and (2) the total value is maximized
[36].

Property 3.1 The MKP is APX-hard and NP-hard [36].

Theorem 1 Problem 1 is equivalent to MKP and is thus NP-hard.

Proof In problem 1, each RB n is analogous to a class. For each RB, it has a maximum
availability of sn = dJ , which is expressed as it can be allocated to a maximum of dJ user.
And the corresponding profit and the volume of the resource allocation item m, n are
wm,n = �rm,n + γ �Fm,n and vm,n = 1, respectively. For each user, it can simultaneously
receive data from a maximum of dV RBs, and thus the total volume should not exceed
V = 1 ∗ dV ∗ M . Thus, problem 2 is equivalent to MKP and is thus NP-hard.  

3.3 Proposed MARL algorithm

In this paper, we employ a low-complexity MARL approach to efficiently address the
optimization problem modeled as a Markov decision process (MDP). The use of MARL,
rather than traditional reinforcement learning (RL), is necessary because the schedul-
ing problem involves allocating multiple RBs to different users. Each agent in the system
independently executes a scheduling policy to assign one RB to the most suitable user.
As Fig. 4 shows, the proposed algorithm comprises a training component and an exe-
cution component, where the execution part follows a classic MDP. An MDP is typi-
cally defined by a tuple (s, a, re, s′ ), where s represents the set of states, a denotes the
set of actions, re is the immediate reward received upon transmitting from one state to
another.

• State: For wireless resource allocation problem, the wireless conditions, especially
instantaneous rate rn,t
ins
and average rate rn,tave
for each agent n, are contained in the
state st . Thus, each node’s state is represented in two dimensions by instantaneous
and average speed, st : (rn,t n,t . Reminding that the update of average rate involves
ins , r ave )

the influence of the previous action at.

• Action: In a single-carrier scheduling scenario, the action involves determining
which UE should be scheduled to occupy the sole RB in each scheduling period (i.e.,
Transmission Time Interval, TTI). In this multi-resource multiuser allocation sce-
nario, each resource is assigned an agent responsible for allocating the corresponding
RB resource. Each  agentcan allocate its assigned RB to at most dJ users. Therefore,
M
the action space is . In For DRL, the prediction value at is designed to be met-
dJ
rics, which determine the wireless resource allocation action for each UE. The BS can
choose the UE with the largest metric as the scheduled UE in each TTI t.
• Reward: The reward design is fundamental in DRL, as it significantly influences both
system performance and training efficiency. In this paper, both the average through-
put and UE fairness are considered into the reward. To quantify the UE fairness,
Jain’s fairness index (JFI) is used. The calculation of JFI is shown in Eq. 3. The details
of reward calculation can be explained as Eq. (12) for each episode.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 12 of 26

Fig. 4 The overall architecture of MARL algorithm for SCMA. Training each AI agent aims to obtain the
optimal neural network parameters to maximize the Q-value. At each scheduling time slot, based on the
current wireless network state and the current neural network state of each AI agent, resource allocation is
performed to obtain the optimal result

ave
ren,t = µm,n,t (�rm,n,t + γ �Fm,n,t ), ∀m ∈ M (12)

where µm,n = 1 indicates that RB n is allocated to user m. The execution part of the algo-
rithm could run continuously to get the wireless resource allocation result every TTI.
Moreover, the algorithm will insert the tuple (s, a, re, s′ ) to the experience learning mem-
ory to support the training tasks.
In the training phase, the algorithm samples a random mini-batch of transitions from
p
the experience learn memory Em. The Q-learning algorithm updates its policy follow-
ing the value iteration function that is proven to converge. It computes the expected
cumulative reward by combining the immediate reward RE and the Q-value function
of the next state. It improves the policy by greedily taking the action that maximizes the
Q-value in the future steps. Based on the centralized action-value function, the gradient
of the training network can be computed using Eq. (13), allowing gradient descent to be
performed on Eq. (14). Moreover, the smoothing factor η is set to 0.9.

ren,t , if episode terminates at step t+1
yn,t = ins , r ave ), a′ ), otherwise
ren,t + η max Q((rn,t (13)
n,t n

ins ave
descentn,t = (yn,t − Q((rn,t , rn,t ), an,t ))2 (14)

here (rn,t n,t is the state of the agent n at step t, an is the best action in the next step,
ins , r ave ) ′

and an,t is the action of the agent n at step t.

Considering the fully cooperative execution task, the optimal neural network is used
to predict the Q-value based on the current wireless network state, and the L-W greedy
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 13 of 26

method is used to achieve the optimal system scheduling result for all agents. The "L-W"
distinguishes weight-updating greedy from the basic greedy algorithm, and it focuses on
the weight variables and efficiency calculations compared with the basic greedy algorithm,
where µm,n = 1 indicates that RB n is allocated to user m. The total execution process is
showed as Algorithms 1, 2 and 3:
Algorithm 1 Multi-Agent Deep Reinforcement Learning Method

Algorithm 2 Procedure 1: Training Network

Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 14 of 26

Algorithm 3 Procedure 2: L-W Greedy method

The complexity of this method focus on the L-W greedy method for each schedul-
ing episode t. In the algorithm 3, The complexity of sorting the actions matrix value
Am,n is O(MN ∗ lg(MN )), and the complexity of allocate µm,n is O(MN), where M
and N are the total number of the clients and resource blocks, respectively. Mean-
while, considering that the algorithm needs to be computed for each RB loop (with
a maximum of N calculations), the overall worst complexity of this method is
O(N ∗ (MN ∗ lg(MN ))) = O(MN 2 ∗ lg(MN )).

3.4 DeepSeek‑enpowered algorithm
In the proposed LLM-based resource allocation strategy, the resource allocation is
determined based on the instantaneous rate conditions of individual users. In this
approach, the instantaneous rate condition of each user is provided as input, and the
corresponding resource allocation outcome is generated as the output by the LLM.
Specifically, these rate conditions are fed into the model as prompts, and the resulting
resource allocation decisions are extracted from the text output of the LLM, as illus-
trated in Fig. 5. Finally, the scheduling module utilizes the proposed results generated by
DeepSeek to allocate wireless resources. Unlike traditional DRL-based methods, which
rely on custom-designed architectures, the LLM-based approach can leverage a general-
purpose language model for resource allocation. Additionally, while conventional DRL
approaches typically process numerical inputs directly, the LLM-based method requires
converting textual prompts into numerical values following a predefined format. If the
model’s output does not conform to this format, accurate resource allocation becomes
unfeasible. In this work, we employ DeepSeek [7] as the underlying LLM.
In our work, we employ two learning-based approaches to address the problem. The
first approach is a semi-automatic method, in which only the core resource allocation
logic is handled by the LLM. The logical relationships among constraints, as well as the
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 15 of 26

Fig. 5 The overall architecture of LLM algorithm for SCMA

rate and fairness gains of RB allocation, are pre-computed. The constraints and optimi-
zation objectives related to RB allocation are provided as inputs to the LLM, and the
resulting RB resource allocation is generated as output.
Algorithm 4 Prompt 1 of SCMA resource allocation

As described in Algorithm 4, the LLM is also required to determine the optimal

resource allocation strategy based on constraints 1, 2, and 3, and output the result in
standard JSON format. During each scheduling epoch, the rate gain and the correspond-
ing fairness gain are calculated and used as inputs to the LLM, which then computes the
RB allocation result. The LLM’s task is to maximize the system objective based on the
preprocessed gain values.
The second approach is fully automatic, where the instantaneous rate conditions of
each user and the latest target requirements are provided as references in the LLM’s
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 16 of 26

prompt. The LLM must automatically compute the resource allocation result according
to the target function Eq. (1) and the given constraints Eq. (2) strictly.
Algorithm 5 Prompt 2 of SCMA resource allocation

As described in Algorithm 5, the LLM is required to determine the optimal resource allo-
cation strategy based on constraints 1, 2, and 3, and output the result in standard JSON
format. During each scheduling epoch, only the instantaneous rate value of one RB per user
and the current average rate of each user are provided as input to the LLM. The LLM then
computes the exponential moving average throughput, fairness value, RB allocation gain for
each user, and the final RB allocation result.
Prompt 2 requires the LLM to handle the complex relationships among the constraints
and the optimization objective and ultimately producing a result. In order to minimize pro-
cessing complexity and and enhance efficiency, we designed Prompt 1 to focus exclusively
on core resource allocation issues.

4 Experimental design
4.1 Experiment setup
To evaluate the performance of the proposed MARL and LLM-based algorithm, we imple-
ment it using a simulator system and Keras tools [37] using NVIDIA GPU A5000, and com-
pare it with the well-known PF scheduling rules [11] and MASR_SCMA method [12]. For
PF scheduling algorithm, UE should be chosen according to
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 17 of 26

ins
rm,t
m∗ = arg max ave . (15)
m∈M rm,t

where the average throughput rm,t

ave
is updated according to
ave ave ins
rm,t = (1 − 1/Tc) ∗ rm,t−1 + Tc ∗ rm,t (16)

with Tc as the window size for averaging, and is set to 10 in this experiment.
We consider a single cell containing multiple clients. A RB consists of several consecu-
tive subcarriers [38]. In our simulation, each subchannel comprises one RB. The simula-
tion parameters are summarized in Table 1. We use the pathloss model as [39].
For the MARL model, the neural networks (NNs) used in each DRL agent are fully
connected ones with 3 hidden layers, the first layer contains 1024 neurons, the second
layer contains 512 neurons, the third layer contains 256 neurons. ReLU function is used
as the activation function for all the hidden layers, similar to [3]. At last, the linear acti-
vation is used for value function. Moreover, we used DeepSeek reasoner model which
is upgraded to DeepSeek-V3.2-Exp, and API access method ([Link]
chat/​compl​etions) in this work.
In the experiment, we compare the PF, MARL, MASR_SCMA, MARL_SCMA and
DeepSeek scheduling algorithms 10 independent times and get similar results. The PF

Table 1 Simulation parameters

Parameter Value

Wireless system simulation parameters

Number of RBs 4
Subcarriers per RB 12
Subcarrier spacing 15 KHz
Bandwidth per RB 180 KHz
BS Tx Power 30 dBm
Pathloss model 128.1 + 37.6*log10 (d[km])
Fading model Rayleigh Fading
Antenna type Omnidirection
Doppler shift 30 Hz
Thermal noise density −174 dBm/Hz
Modulation/coding rate settings M-QAM
Number of users 6
Traffic models Best effort
MARL parameters
Optimizer RMSprop
Learning rate 0.01
Exploration rate 0.001
Batch size 256
Target network update frequency 300
Replay memory capacity 8000
DeepSeek parameters
Temperature 0
Max tokens 3000
Model DeepSeek-V3.2-Exp
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 18 of 26

and MARL algorithms primarily address resource scheduling, where each RB can be
allocated to only one user. In contrast, MASR_SCMA, MARL_SCMA_3_2 and MARL_
SCMA_2_1 focus on SCMA resource scheduling, which allows an RB to be shared
among multiple users for data transmission. MARL_SCMA_3_2 and MARL_SCMA_2_1
 and (2, 1), respectively. Additionally, dV is set to 2,
correspond to the (L, W)pairs (3, 2)
N −1
dJ is calculated as dJ = = 3, where the Number of RBs N = 4 . Thus, the
dV − 1
) pairs
( d J , dV   are set to (3, 2). The best number of users is calculated as
N
M∗ = = 6. We assume a cellular network with 6 users, three of whom are
dV
mobile with speeds of 3, 5, and 10 meters per second, respectively, while the remaining
users are stationary.

4.2 Evaluation metrics
We first define the evaluation metrics to assess transmission performance: (1) Average
Bitrate: This metric evaluates the efficiency of the scheduling by measuring the aver-
age bitrate. (2) Fairness: We use Jain’s fairness index to compare the variance among all
users.

5 Results and discussion

5.1 Performance comparison for the MARL method
We compared the fairness of various DRL-based algorithms. As shown in Fig. 6, fairness
improves with increasing parameter values, indicating that the weight of fairness in the
reward function becomes more prominent, leading the algorithms to pay more attention
to the effectiveness of fairness. When the parameter exceeds 20, the fairness of MARL,
MARL_SCMA_3_2, and MARL_SCMA_2_1 all exceed 0.99.
With the parameter γ set to 1, 10, and 100, we compared the average rate and the
cumulative distribution function (CDF) of user rates across different algorithms. To bet-
ter illustrate the learning process, a focused comparison was conducted between the
PF and MARL algorithms. As seen in Fig. 7c, when γ = 1, the MARL algorithm prior-
itizes users with higher data rates, resulting in few resource is allocated to users under
unfavorable channel conditions. Meanwhile, users with better channel conditions are

Fig. 6 Jain’s fairness values using MARL, MARL_SCMA_3_2 and MARL_SCMA_2_1 method with different
numbers of parameter γ
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 19 of 26

Fig. 7 Average bitrates and CDF of PF, MARL, MASR_SCMA, MARL_SCMA_3_2 and MARL_SCMA_2_1
methods for the parameter γ = 1: a, b is the average bitrate value. c is the CDF of average bitrate. Results
for the parameter γ = 10: d, e is the average bitrate value. f is the CDF of average bitrate. Results for the
parameter γ = 100: g, h is the average bitrate values. i is the CDF of average bitrate

allocated more wireless resources, leading to a higher overall average throughput. Fur-
thermore, since SCMA enables multiple users to share wireless resources, the system
throughput of MASR_SCMA, MARL_SCMA_3_2 and MARL_SCMA_2_1 exceeds that
of both the PF and standard MARL scheduling algorithms. The MASR_SCMA has the
best performance of average bitrate and the worst performance of fairness, because it
only focuses on the average bitrate. When γ is increased to 100, as shown in Fig. 7i, the
fairness performance of the MARL algorithm improves significantly, with a noticeable
reduction in disparity among user average rates. At the same time, the average rate of
the MARL algorithm decreases by approximately 0.03 Mbps, as shown in Fig. 7d and g.
In contrast, the average rate of MARL_SCMA remains stable or even improves, while
fairness is enhanced.

5.2 Performance comparison for the DeepSeek based method

We compared the performance of different DeepSeek prompts. As shown in Fig. 8, fair-
ness improves with an increase in the parameter γ , indicating that the weight of fair-
ness in the reward becomes more significant, and the algorithms place greater emphasis
on achieving fairness. The fairness value of Prompt 1 approaches 1.0 as the parameter
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 20 of 26

Fig. 8 Jain’s fairness values using DeepSeek_prompt_1 and DeepSeek_prompt_2 methods with different
numbers of parameter γ

Fig. 9 Average bitrates and CDF of different deepseek prompts for the parameter γ = 1: a is the average
bitrate values. d is the CDF of average bitrate. Results for the parameter γ = 10: b is the average bitrate
values. e is the CDF of average bitrate. Results for the parameter γ = 100: c is the average bitrate values. f is
the CDF of average bitrate

γ increases, demonstrating its quick responsiveness to parameter changes. Prompt 2

achieves a fairness value exceeding 0.994.
With the parameter γ set to 1, 10, and 100, we also compared the rate, the CDF of
the rates, and the average rate for different users across various prompts. As shown in
the rate figures, the average rate of deepseek approaches 2 Mbps. From Fig. 9a–c, it can
be observed that as the number of training steps increases, the average rates for both
prompts become similar. As seen in Fig. 9d–f, Prompt 1 performs better than Prompt 2
by considering the rate and fairness, particularly when the parameter γ is set to 100.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 21 of 26

5.3 Performance comparison for all methods

Finally, we compared the performance of PF, MARL_SCMA_3_2 (referred to as MARL_
SCMA in the following figures), and deepseek prompt 1 (referred to as deepseek in the
following figures), since MARL_SCMA and deepseek prompt 1 demonstrated better
performance compared to other methods. As shown in Fig. 10, the fairness improves
with increasing values of the parameter γ for deepseek prompt 1. Moreover, the fairness
results for both methods exceed 0.98 .
With the parameter γ set to 1, 10, and 100, we also compared the rate, the CDF of
the rates, and the average rate across different users for the three methods. As shown
in the rate figures, the rates of MASR_SCMA, MARL_SCMA and deepseek are signifi-
cantly higher than that of the PF method. This is because the SCMA method provides a
higher rate gain, and the average rate is about 2 Mbps using these three methods. From
Fig. 11d–f, the fairness value of deepseek prompt 1 is higher than the others. In terms of
fairness, The deepseek prompt 1 demonstrates a faster response to changes.
To demonstrate the stability of average rate and fairness values across multiple simu-
lation runs, we compiled the mean values and standard deviations of average rate and
fairness for the PF, MASR_SCMA, MARL_SCMA, and DeepSeek method as shown in
Fig. 12a and b. It can be observed that the standard deviations of the MARL_SCMA
and DeepSeek methods are relatively larger compared to those of PF and MASR_SCMA,
indicating greater variability in their resource allocation results. However, the overall
variation trend is not significant, suggesting that AI-based methods exhibit more notice-
able changes in resource allocation outcomes.
In order to explore performance bound and scalability, we also focus on the average
rate of PF, MASR_SCMA, MARL_SCMA and DeepSeek methods under diverse user
scenarios as Fig. 13 shows. The number of RBs is 6 and the number of users increases
from 2 to 14 in this scenario. It could be seen that average data rate decreases as the
number of users increases for all methods. When the number of users is 2, the average
bitrate is about 2Mbps for all methods. The average bit rate gradually decreases begin-
ning with 6 users for PF method. The average bitrate of MASR_SCMA, MARL_SCMA

Fig. 10 Jain’s fairness values using MARL_SCMA and DeepSeek methods with different numbers of
parameter γ
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 22 of 26

Fig. 11 Average bitrates and CDF of PF, MASR_SCMA, MARL_SCMA and DeepSeek methods for the
parameter γ = 1: a is the average bitrate values. d is the CDF of average bitrate. Results for the parameter
γ = 10: b is the average bitrate values. e is the CDF of average bitrate. Results for the parameter γ = 100: c is
the average bitrate values. f is the CDF of average bitrate

Fig. 12 Mean values and standard deviations of average rate and fairness for PF, MASR_SCMA, MARL_SCMA
and DeepSeek methods: a is the mean and standard deviation of average rate values. b is the mean and
standard deviation of fairness values

and DeepSeek begin to gradually decrease to about 1.3 Mbps when the number of users
is 10. They have better performance compared with PF method because of SCMA multi-
access gain and multiuser diversity gain. When the number of users increases to 14, the
average bitrate is about 0.3 Mbps, 1.0 Mbps, 0.9 Mbps, 0.9 Mbps for the PF, MASR_
SCMA, MARL_SCMA and DeepSeek methods correspondingly.

6 Challenges and open questions

Although AI-based methods presents a promising approach to enable intelligent
resource allocation decision-making in complex and dynamic wireless environments,
it also requires to overcome several key challenges. In the following, we discuss the
research challenges and open questions associated with the future use of AI-based
methods for resource allocation.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 23 of 26

Fig. 13 Average rate of PF, MASR_SCMA, MARL_SCMA and DeepSeek methods under diverse user scenarios

6.1 Latency and cost

As discussed in [28, 40], In general, the size of LLMs is extremely large due to their
large number of parameters. DeepSeek, for example, has 1.5-671 billion parameters.
As a result, running LLMs requires significant computation resources, often neces-
sitating the use of cloud-based services, which can introduce long latency. Moreover,
to leverage the reasoning capabilities of LLMs for resource allocation analysis, the
required time is also increasing. Although current generalized LLM inference laten-
cies typically range from tens to hundreds of seconds, researchers [41, 42] are now
working to significantly reduce this latency to the order of seconds by optimizing dis-
tributed GPU service frameworks and service rescheduling mechanisms. To address
the issues of long latency and computation time, the development of small LLMs tai-
lored for resource allocation is also essential [28]. With further optimization of LLM
services in edge networks, when inference latency is generally reduced to the second
and millisecond levels, LLM models can be utilized for hierarchical and tiered wire-
less resource scheduling across different time granularities.
In terms of cost, with the large-scale adoption of open-source large models such as
DeepSeek, researchers can utilize these models for training and inference, significantly
reducing expenses—even to zero.

6.2 Vulnerability
LLMs are highly sensitive to prompts. Inappropriate prompts may lead to responses in
incorrect formats or with non-compliant content. Therefore, during the training pro-
cess, it is necessary to integrate prompt engineering. By detecting and updating prompt
requirements, the correctness of the feedback results can be enhanced. Additionally,
methods such as redundant pre-sending and deduplication can be employed to improve
processing accuracy.
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 24 of 26

6.3 Privacy
Preserving the privacy of the users is the primary concern of mobile and service provid-
ers. The privacy concerns demand significant attention. LLMs are trained on massive
datasets and inadvertently memorize and regurgitate sensitive information from their
training data. Once they encounter adversarial attacks, they are quite likely to reveal the
sensitive information. Moreover, the complexity and lack of transparency in how LLMs
process and generate responses make it difficult to control and predict what information
they may disclose, thereby further obscuring privacy risks. To improve the model trans-
parency, we can develop explainable AI tools or white-box LLMs to provide insights into
the decision-making process, thereby improving its predictability and reliability. Alter-
natively, one might consider how to train models on datasets belonging to users without
sharing their input data or compromising the security of their personal information.

6.4 Security
The security of deep learning models itself in another challenge, as neural networks
are prune to adversarial attacks. Attackers can affect the training process by injecting
fake training datasets; such injection can lower the accuracy of the models and yield
wrong design, which may affect the network performance. Research in the security of
deep learning or machine learning, in general, remains shallow. To mitigate the poten-
tial harmful training data, robust data validation and sanitization techniques can be
employed to detect and remove anomalous or suspicious data before being incorporated
into the training dataset. Alternatively, adversarial training methods can be employed to
expose the model to adversarial examples during the training phase.

7 Conclusion
This paper formulates the optimization problem for sparse code multiple access in 5G
cellular networks and solves it using multi-agent reinforcement learning and large lan-
guage model-based approaches. Experimental results show that the proposed meth-
ods outperform existing resource allocation schemes, mainly due to the code-division
resource multiplexing mechanism of SCMA, and the advantages of intelligent algorithms
in perceiving and predicting environmental changes while enabling flexible adjustments.
In future work, we plan to optimize this problem in more realistic scenarios, such as pri-
oritizing how to reduce the inference latency of LLM-based resource scheduling meth-
ods by optimizing the LLM service architecture and network environment, or training
more specialized resource scheduling LLMs, thereby enhancing the practicality of the
approach. Additionally, more accurate models for 5G SCMA transmission characteris-
tics and multi-cell massive user scenarios can be established. The robustness of the pro-
posed methods in rapidly changing wireless mobile communication environments will
also be investigated.

Abbreviations
SCMA Sparse code multiple access
MARL Multi-agent deep reinforcement learning
LLM Large language model
eMBB Enhanced mobile broadband
URLLC Ultra-reliable low-latency communications
mMTC Massive machine-type communications
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 25 of 26

OFDMA Orthogonal frequency-division multiple access

UEs User equipment
DRL Deep reinforcement learning
RBG Resource block group
PF Proportional fair
UDNs Ultra-dense networks
CSI Channel state information
BS Base station
BER Bit error rate
MIMO Massive multiple-input multiple-output
B5G Beyond-5G
CDMA Code-division multiple access

Acknowledgements
The authors thank the anonymous reviewers for their suggestions that have significantly enhanced the quality and
presentation of this paper.

Author contributions
Mincheng Zhao designed the scheme and algorithm and organized experiments. Mingqing Han provided SCMA
modeling methods and conducted MARL algorithm validation. Jiyuan Pan conducted the experiments, including the
implementation of the base algorithm. Ligang Du conducted the experiments, including data aggregation and process-
ing. Yongfeng Yan reviewed and revised the manuscript. Fuqiang Li reviewed and revised the manuscript, providing
access to experimental facilities.

Funding
Not applicable.

Data availability
Code was available at GitHub, [Link]

Declarations
Competing interest
The authors declare that they have no conflict of interest.

Received: 10 November 2025 Accepted: 4 March 2026

References
1. H. Nikopour, H. Baligh, in IEEE Proc. PIMRC. Sparse code multiple access (2013)
2. Z. Yuan, G. Yu, W. Li, Y. Yuan, J. Xu, in IEEE Proc. VTC. Multi-User Shared Access for Internet of Things (2016)
3. J. Wang, C. Xu, Y. Huangfu, R. Li, Y. Ge, J. Wang, in IEEE Proc. WCSP. Deep Reinforcement Learning for Scheduling in
Cellular Networks (2019)
4. C. Xu, J. Wang, T. Yu, C. Kong, Y. Huangfu, R. Li, Y. Ge, J. Wang, in IEEE Proc. WCNC. Buffer-aware Wireless Scheduling
based on Deep Reinforcement Learning (2020)
5. C. Kai, X. Meng, L. Mei, W. Huang, Multi-agent reinforcement learning based joint uplink-downlink subcarrier assign-
ment and power allocation for D2D underlay networks. Wireless Netw. 29, 891–907 (2022)
6. ChatGPT: A glimpse of the future. [Link]
7. deepseek Into the unknown. [Link]
8. J. Shao, J. Tong, Q. Wu, W. Guo, Z. Li, Z. Lin, J. Zhang, WirelessLLM: Empowering Large Language Models Towards
Wireless Intelligence (2024) arXiv:​2405.​17053 [[Link]]
9. M. Zhao, M. Han, L. Du, J. Pan, Y. Yan, F. Li, in IEEE Proc. WCSP. AI-based Resource Scheduling in 5G NOMA Cellular
Networks (2025)
10. P. McEnroe, S. Wang, M. Liyanage, in IEEE Proc. ITSC. Towards Latency Efficient DRL Inference: Improving UAV Obstacle
Avoidance at the Edge Through Model Compression, pp. 4242–4249 (2024)
11. G. Song, e.a., in IIEEE Proc. GLOBECOM. Adaptive resource allocation based on utility optimization in OFDM. (2003)
12. M. Cheraghy, W. Chen, in IEEE Proc. ICCC​. Resource Allocation to Maximize the Average Sum Rate of the Uplink SCMA
Networks. (2021)
13. Y. Zheng, J. Cui, X. Lei, Z. Ding, P. Fan, D. Chen, Impact of factor graph on average sum rate for uplink sparse code
multiple access systems. IEEE Access 4, 6585–6590 (2016)
14. M. Zeng, N.-P. Nguyen, O.A. Dobre, Z. Ding, H.V. Poor, Spectral- and energy-efficient resource allocation for multi-
carrier uplink NOMA systems. IEEE Trans. Veh. Technol. 68(9), 9293–9296 (2019)
15. S. Jaber, W. Chen, in IEEE Proc. ICCT​. Subcarrier assignment and power allocation for SCMA energy efficiency (2020)
16. X. Wang, Y. Wang, Q. Cui, K.-C. Chen, W. Ni, Machine learning enables radio resource allocation in the downlink of
ultra-low latency vehicular networks. IEEE Access 10, 44710–44723 (2022)
17. F.D. Rango, N. Cordeschi, F. Ritacco, in IEEE Proc. CCNC. Applying Q-learning approach to CSMA scheme to dynami-
cally tune the contention probability, pp. 1–4 (2021)
18. H.A. Akyildiz, O.F. Gemici, I. Hokelek, H.A. Cirpan, Hierarchical reinforcement learning based resource allocation for
RAN slicing. IEEE Access 12, 75818–75831 (2024)
Zhao et al. J. Wirel. Commun. Netw. (2026) 2026:39 Page 26 of 26

19. S. Bhardwaj, D.-S. Kim, Deep Q-learning based sparse code multiple access for ultra reliable low latency communica-
tion in industrial wireless networks. Telecommun. Syst. 83(4), 409–421 (2023)
20. X. Du, T. Wang, Q. Feng, C. Ye, T. Tao, L. Wang, Y. Shi, M. Chen, Multi-agent reinforcement learning for dynamic
resource management in 6G in-X Subnetworks. IEEE Trans. Wireless Commun. 22(3), 1900–1914 (2023)
21. D.C. Bikkasani, M.R. Yerabolu, AI-driven 5G network optimization: a comprehensive review of resource allocation,
traffic management, and dynamic network slicing. Am. J. Artif. Intell. 8(2), 55–62 (2024)
22. M. Dubey, A.K. Singh, R. Mishra, AI based resource management for 5G network slicing: history, use cases, and
research directions. Concurr. Comput.: Pract. Exp. 37(2), 66 (2024)
23. N. Kan, J. Zou, K. Tang, C. Li, N. Liu, H. Xiong, in IEEE Proc. ICASSP. Deep Reinforcement Learning-based Rate Adapta-
tion for Adaptive 360-Degree Video Streaming, pp. 4030–4034 (2019)
24. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, I. Polosukhin, in ACM Proc. NIPS. Atten-
tion is all you need (2017)
25. D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X.B. al, DeepSeek-R1: Incentivizing Reason-
ing Capability in LLMs via Reinforcement Learning (2025) arXiv:​2501.​12948 [[Link]]
26. A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, e.a. Chenggang Zhao, Deepseek-v3 technical report (2025) arXiv:​2412.​
19437 [[Link]]
27. H. Zhou, C. Hu, D. Yuan, Y. Yuan, D. Wu, X. Liu, C. Zhang, Large Language Model (LLM)-enabled In-context Learning
for Wireless Network Optimization: A Case Study of Power Control (2025) arXiv:​2408.​00214 [[Link]]
28. W. Lee, J. Park, LLM-Empowered Resource Allocation in Wireless Communications Systems (2024) arXiv:​2408.​02944
[[Link]]
29. H. Noh, B. Shim, H.J. Yang, Adaptive Resource allocation optimization using large language models in dynamic wire-
less environments. IEEE Trans. Veh. Technol. 66, 1–6 (2025)
30. L. Bao, S. Yun, J. Lee, T.Q.S. Quek, LLM-hRIC: LLM-empowered Hierarchical RAN Intelligent Control for O-RAN (2025)
arXiv:​2504.​18062 [[Link]]
31. X. Peng, Y. Liu, Y. Cang, C. Cao, M. Chen, LLM-OptiRA: LLM-Driven Optimization of Resource Allocation for Non-
Convex Problems in Wireless Communications (2025) arXiv:​2505.​02091 [[Link]]
32. B. Di, L. Song, Y. Li, in IEEE Proc. ICC . Radio resource allocation for uplink sparse code multiple access (SCMA) net-
works using matching game (2016)
33. A.J. Goldsmith, S.-G. Chua, Variable-rate variable-power M-QAM for fading channels. IEEE Trans. Commun. 45(10),
1218–1230 (1997)
34. F.R. Kschischang, B.J. Frey, H.-A. Loeliger, Factor graphs and the sum-product algorithm. IEEE Trans. Inf. Theory 47(2),
498–519 (2001)
35. H. Zhang, S. Han, W.-X. Meng, in IEEE Proc. VTC-Fall. Multi-Stage Message Passing Algorithm for SCMA Downlink
Receiver, pp. 1–5 (2016)
36. C. Chekuri, S. Khanna, in ACM-SIAM Symposium on Discrete Algorithms. A PTAS for the multiple knapsack problem
(2000)
37. Keras: Deep Learning for humans. [Link]
38. Physical layer aspects for evolved universal terrestrial radio access (utra). 3GPP TS 25.814 bf7.1.0 (2006)
39. H. Nikopour, E. Yi, A. Bayesteh, K. Au, M. Hawryluck, H. Baligh, J. Ma, in IEEE Proc. GLOBECOM. SCMA for downlink
multiple access of 5G wireless networks (2014)
40. H. Zou, Q. Zhao, L. Bariah, M.D. Mehdi Bennis, Wireless Multi-Agent Generative AI: From Connected Intelligence to
Collective Intelligence (2023) arXiv:​2307.​02757 [[Link]]
41. B. Sun, Z. Huang, H. Zhao, W. Xiao, X. Zhang, Y. Li, W. Lin, in USENIX Symposium on OSDI. Llumnix: Dynamic Scheduling
for Large Language Model Serving (2024)
42. Y. Fu, L. Xue, Y. Huang, A.-O. Brabete, D. Ustiugov, Y. Patel, L. Mai, in USENIX Symposium on OSDI. ServerlessLLM: Low-
Latency Serverless Inference for Large Language Models (2024)

Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.