Learning-Based Collaborative MEC for LLM Inference with Soft-Deadline Awareness via Transformer-Enhanced PPO

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

1

Learning-Based Collaborative MEC for LLM Inference with Soft-Deadline Awareness via Transformer-Enhanced PPO Ngoc Hung Nguyen and Bjorn Landfeldt Department of Electrical and Information Technology, Lund University, P.O. Box 117, Lund 22100, Sweden Corresponding author: Ngoc Hung Nguyen, email: ngoc hung.nguyen@eit.lth.se

arXiv:2608.02031v1 [cs.DC] 3 Aug 2026

Abstract—This paper investigates collaborative mobile edge computing (MEC) servers for large language model (LLM) inference under soft deadline constraints. In this system, to improve the quality of service, computations are expected to be completed within their deadlines. However, due to dependencies among tasks or subtasks, any missed deadline can lead to catastrophic consequences for the entire request. In this context, this work proposes an extended deadline mechanism with constrained flexibility. The main challenges lie in handling large-scale computations under strict latency constraints while limiting the number of allowable deadline extensions, especially in the presence of task dependencies within each request. To tackle these challenges, we develop a transformer-enhanced proximal policy optimization (PPO) framework that enables efficient collaboration among MEC servers. The proposed approach aims to maximize the number of tasks completed within their deadlines while minimizing the use of deadline extensions. By capturing temporal dependencies and cross-server interactions, the transformer improves decisionmaking for task migration. Simulation results demonstrate that the proposed method significantly outperforms conventional PPO and heuristic-based approaches in terms of task completion rate and overall system efficiency. Index Terms—Mobile edge computing, collaborative MEC, task migration, LLM inference, transformer, PPO.

I. I NTRODUCTION

Recently, the rapid development of artificial intelligence (AI) has changed the way people use applications. Instead of using normal search engines, people increasingly tend to raise questions using generative AI (GenAI) applications. The applications not only provide an information summarization but can also process many other types of tasks, such as video creation, image analysis, etc. These services are attractive to a large number of customers but comes at the cost of increased computation for each request or prompt, bringing the system to a high workload state, with large response times. However, the potential application of this AI architecture is wideranging in the future, including real-time support (e.g., Avatarrecommendation system) in Virtual and Augmented Reality (XR) systems, or orchestration of intelligent transportation systems in smart cities [1], [2]. Thus, research is ongoing on how the computation of these kinds of tasks gradually becomes more and more crucial in accuracy, latency, and energy consumption [3], [4]. Mobile edge computing (MEC) provides services close to the edge of the computing system, near end users. However, MEC has limited computational resources compared with central cloud platforms. Nevertheless, MEC has an advantage over

cloud platforms in that latencies can be reduced and network bottlenecks can be avoided [5]. However, computing highworkload tasks with complex architectures, such as LLMs, is not easy for a single MEC server. Instead of using a single server to process heavy workloads, many works have proposed collaboration mechanisms among MEC servers. This is expected to become an important direction for improving task computation efficiency in MEC systems [6], [7]. In fact, once heavy workload applications such as LLMs can be successfully deployed on edge infrastructure, they can enable real-time intelligent services with low latency and enhanced privacy for end devices. In practice, the communication load in both uplink and downlink can be quite heavy, such as video, raw images, or data files. Computation at the edge is expected to reduce latency and improve users’ Quality of Service (QoS) experience overall [8]. Therefore, enabling efficient heavy workload execution on MEC systems is not only beneficial but also necessary for supporting next-generation intelligent applications. Therefore, in this work, we utilize an edgeedge collaboration mechanism to enhance the performance of computation for heavy workload tasks, especially for LLM models [9]. Deep reinforcement learning (DRL) has been widely adopted for solving complex decision-making problems, with various advanced algorithms developed in recent years. Among them, proximal policy optimization (PPO) has emerged as a robust and stable policy-gradient method for many practical applications [10]. However, PPO may not consistently achieve high performance in highly dynamic environments due to its sensitivity to hyperparameter settings and its limited capability in capturing long-term temporal dependencies. In the considered MEC system, task migration decisions are influenced not only by the current system state but also by historical information, such as queue evolution, workload fluctuations, and task dependency patterns. Conventional PPO, which relies primarily on instantaneous observations, may lead to myopic decisions under such conditions [11], [12]. To address this limitation, we incorporate a transformer architecture into the PPO framework. By leveraging self-attention mechanisms, the transformer effectively captures temporal dependencies and cross-server interactions from historical system states, enabling more informed and context-aware decision-making for task migration in collaborative MEC systems. 1) We propose a collaborative MEC framework for LLM inference, where each request is modeled as a hierarchi-

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

𝑇!

(1 )

𝑇$

,!

,!

) (1 𝑇#,!

(3)

𝑇%,!

𝑇&,! 𝑅!

(2)

,!

UE 1

(.)

: Task computing request : Offer to user : Interdependence relative

: Create groups of tasks request : Result of computing

Fig. 1. Illustration of system model.

cal directed acyclic graph (DAG) consisting of multiple tasks and subtasks with dependency constraints. Based on this framework, we formulate a task migration and soft-deadline-aware scheduling problem, and identify its key challenges arising from task dependencies, resource heterogeneity, and latency constraints. The problem is then modeled as a Markov decision process (MDP), capturing system dynamics including queue states, task dependencies, and communication conditions as well as servers’ information. 2) We develop a transformer-enhanced PPO algorithm that exploits temporal dependencies and cross-server interactions to improve decision-making in task migration. We design a normalized system-level reward function that aligns with the optimization objective, enabling stable and effective policy learning. 3) Simulation results demonstrate that the proposed approach significantly outperforms conventional PPO and static (no migration) task baselines in terms of task completion rate and deadline efficiency. II. S YSTEM M ODEL AND P ROBLEM F ORMULATION A. System Model As illustrated in Fig. 1, we consider an edge-edge collaboration system expected to enhance the computation capacity of high-demand tasks. The system includes a set of |S| MEC servers S = {S1 , S2 , · · · , S|S| } and a set of user U = {u1 , u2 , · · · u|U | }. Each of these servers is provided with a computation capacity of fs [cycles/second]. Note that the system is heterogeneous, meaning that the computational resources of each MEC may vary. Each MEC is distributed in a different region within the deployment area, and each region is serviced by a base station using orthogonal frequency multiple access (OFDMA). The designed system allows for server collaboration through a high-speed physical backhaul by applying remote direct memory access (RDMA). Therefore, tasks in different

2

MEC nodes can be migrated or shared with ultra-low communication latency, enabling real-time cooperative processing across the entire network. Collaboration among MEC servers plays a crucial role in this work, as the system is expected to [sub] handle delay-sensitive tasks modeled by DAGs. Let Tk,i = (1)

(2)

(d

)

{Tk,i , Tk,i , . . . , Tk,ik,i } denote the set of subtasks associated with task Tk,i in request Ri , where requests arrive according to a Poisson process with rate λ [requests/user/s], and dk,i denotes the number  of subtasks in Tk,i . Each  task is represented as [sub] Tk,i ≜ Tk,i , Dk,i , αk,i , βk,i , γk,i , where k ∈ {1, . . . , Ki } and Ki denotes the number of tasks in request Ri . Accordingly,   i request Ri is represented as Ri ≜ {Tk,i }K k=1 , di , ri , ai , where di and ri [bits] denote the upload and download data sizes of request i, respectively, and ai denotes the time at which the request arrives in the system. Moreover, Dk,i denotes the [sub] set of dependency relationships among the subtasks in Tk,i , i.e., the edges of thecorresponding DAG. For each subtask (j) (j) (j) (j) (j) (j) j, we define Tk,i ≜ αk,i , βk,i , γk,i , where αk,i , βk,i , and (j)

γk,i denote the required number of CPU cycles, the deadline, and the number of deadline-extension requests of subtask j, respectively. Accordingly, the total computational requirement Pdk,i (j) of task Tk,i is αk,i = j=1 αk,i , its deadline is given by (j) βk,i = maxj∈{1,...,dk,i } βk,i , and its total number of deadlinePdk,i (j) extension requests is γk,i = j=1 γk,i . In this work, we aim to maximize the number of tasks completed within their deadlines while minimizing the number of deadline-extension requests. To enhance computational performance, MEC collaboration is conducted as follows: (1) the host MEC server selects a subset of tasks that require assistance for execution, (2) the host MEC collects system information from neighboring MEC servers to determine appropriate helper servers, (3) the selected helper MEC servers execute the migrated tasks and return the computed results to the host MEC server, and (4) the host MEC server waits for the completion of all tasks within a request, aggregates the results, and delivers the final outcome to the user. Note that the system allows deadline extension, where each extension increases the task deadline by an amount equal to its computation time, excluding queuing delay. To mitigate excessive queuing delays, each task is subject to a limited number of deadline extensions for all belonging subtasks. B. Problem Formulation Each user is assumed to move according to a Markovian mobility process, following the model in [13]. Let the position of user u ∈ [1, U ] at time t be denoted by (xu (t), yu (t)) ∈ A, where A [m2 ] represents the considered area. The position of base station s ∈ [1, |S|] is denoted by (xs , ys ) ∈ A. Accordingly, the distance between p user u and base station s at time t is given by lu,s (t) = (xu (t) − xs )2 + (yu (t) − ys )2 . The channel coefficient between user u and resource block n ∈ [1, N ] of base station s at time t is expressed as gu,s,n (t) hu,s,n (t) = lu,s , where p ∈ [2, 4] is the path-loss exponent (t)p/2 and gu,s,n (t) ∼ CN (0, 1) denotes small-scale fading.

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

Let xu,s,n (t) ∈ {0, 1} denote the OFDMA resource allocation variable, where xu,s,n (t) = 1 if resource block n of base station s is assigned to user u at time t, and xu,s,n (t) = 0 otherwise. Due to the orthogonality of OFDMA, each resource block can be allocated to at most one user at a given time, i.e., U X

xu,s,n (t) ≤ 1,

∀s, n, t.

(1)

u=1

Moreover, each user can be assigned at most one resource block at each time slot, i.e., |S| N X X

xu,s,n (t) ≤ 1,

∀u, t.

(2)

s=1 n=1

The achievable data rate of user u on resource block n of base  station s at 2 time t is given by Ru,s,n (t) = P |h (t)| , where Bn denotes the bandwidth Bn log2 1 + u Nu,s,n 0 Bn of each resource block, Pu is the transmit power of user u, and N0 [W/Hz] denotes the additive white Gaussian noise power spectral density. Accordingly, the noise power over each resource block is given by N0 Bn . The total data rate of user u associated with PN base station s at time t is expressed as Ru,s (t) = n=1 xu,s,n (t)Ru,s,n (t). Let σi = u indicate that request Ri is generated by user u. Then, the total transmission delay associated with request Ri is given by [comm] i +ri with u = σi . Note that, with notation τi (t) = Rdu,s (t) , xu,s (t) ∈ {0, 1}, it implies that s ∈ [1, |S|] is selected as the host MEC server for computing task Ri from user u if xu,s (t) = 1 . Thus, only one MEC should be selected as the host MEC, and the resource block should be in the associated base station then: |S| X

xu,s (t) = 1, s.t xu,s,n (t) ≤ xu,s (t),

∀u, s, n, t

(j)

the final position of subtask Tk,i in queue qs,m∗ . Then, the (j) computation delay of subtask Tk,i generated by user σi = u is

∗ P ∗ α q (p) s,m [comp],(j) p , where α qs,m∗ (p) given by τk,i = p=1 fs,m∗ denotes the required number of CPU cycles of the subtask (j) at position p in the queue. Thus, Tk,i satisfies deadline if: [comm] [comp],(j) (j) (j) ai + τi (t) + τk,i (t) ≤ βk,i (t) + (1 − φk,i )Y , here, the deadline can be changed if scheduling fails, thus it depends (j) on the time. Y is a number which is larger than any βk,i (j) and φk,i ∈ {0, 1} denotes a binary variable indicating whether (j) subtask Tk,i is completed before its deadline. For the computation dependency condition, a task Tk,i is ready for execution only if all its preceding tasks have been completed. Define φk,i (t) ∈ {0, 1} as a binary variable indicating whether task Tk,i has been completed, where φk,i (t) = 1 if it is completed and 0 otherwise. Then, the readiness condition of task Tk,i is given by Y φk′ ,i (t) = 1. (5) Tk′ ,i ∈Dk,i

The task Tk,i is considered successfully completed if and only if all its subtasks are completed before their deadlines, i.e., dk,i X

fs,m = fs ,

(4)

m=1

where fs,m denotes the computation capacity of resource m at server s. Each computing resource maintains an independent queue for subtask processing, and all queues operate under the earliest-deadline-first scheduling and non-preemptive policy. At each decision epoch, the system observes all MEC servers and their available computing resources. For load balancing, we assume that once an MEC server s is selected to process a task, its subtasks are assigned to the computing resource with the shortest queue within that server. Let m∗ = arg minm∈[1,M ] |qs,m | denote the computing resource with the shortest queue at server s, whose queue is represented by qs,m∗ = {qs,m∗ (1), qs,m∗ (2), . . . , qs,m∗ (|qs,m∗ |)}. Since the queue is dynamically updated as new subtasks are assigned, the position of a subtask may change over time. Let p∗ denote

(6)

otherwise, the task is considered failed. If a subtask is predicted to miss its deadline during scheduling, the system allows a (j) deadline extension, which is recorded by γk,i (t). In this work, the number of deadline extensions is limited by an upper bound, expressed as dk,i X

s=1

M X

(j)

φk,i (t) = dk,i ,

j=1

(3)

For computation, we assume that each MEC server is equipped with M computing resources such that

3

(j)

γk,i (t) ≤ Γmax k,i ,

(7)

j=1

here, Γmax k,i is the maximum number of deadline extension for task Tk,i . Based on the above system model, we aim to jointly optimize task assignment and deadline extension decisions to maximize the number of successfully completed tasks while minimizing the number of deadline extension requests. Let xk,i,s (t) ∈ {0, 1} denote the task migration variable, where xk,i,s (t) = 1 if task Tk,i is processed at MEC server s at time t, and 0 otherwise. Each task is assigned to exactly one MEC server, i.e., |S| X

xk,i,s (t) = 1,

∀k, i, t,

(8)

s=1

which implies that the task can be executed at any MEC among |S|, including the host MEC. Therefore, the optimization problem is formulated where η is a weighting coefficient that controls the importance of minimizing the number of deadline extensions. Let x =

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

Transformer-enhanced PPO Policy Network Score each MEC server (context-aware future) Transformer Encoder Temporal Dependency Learning Cross-Server Correlation Global Context Extraction

Decision Making Action: Select Helper MEC Server

Value Network Estimate long-term reward

Advantage estimation

State Sequences

MEC 1 Information

Host MEC Migratable Tasks

Reward

MEC 2 Information

Next State

MEC N Information

Fig. 2. Transformer-enhanced PPO framework for contextaware MEC server selection.

{xu,s,n (t), xu,s (t), xk,i,s (t) | ∀u, k, i, s, n, t} denote the collection of all binary decision variables, thus the problem formulation can be expressed by:

P1 :

max x

s.t.

XX i

φk,i (t) − η

dk,i XXX i

k

k

(j)

γk,i (t)

(9)

j=1

(1), (2), (3), (4), (5), (6), (7), (8),

(9a)

(j) φk,i (t), φk,i (t) ∈ {0, 1}, (j) γk,i (t) ∈ Z+ , ∀k, i, j, t.

(9b)

(j)

∀k, i, j, t,

(9c)

(j)

Note that φk,i (t), φk,i (t), and γk,i (t) are outcome variables determined by the task-assignment decision x and the system dynamics. Therefore, constraints involving these variables implicitly restrict the feasible set of x. In this work, we focus on the optimization of task migration; thus, xu,s,n (t) and xu,s (t) are determined by the predefined random resource-allocation and greedy association policies, respectively, whereas xk,i,s (t) is optimized by the proposed learning agent. III. P ROPOSED M ETHOD This work focuses on improving the efficiency of collaboration among MEC servers. To simplify the system model, each user is associated with a host MEC server based on a greedy distance-based policy, i.e., the nearest server is selected for task offloading. We assume that each MEC server can broadcast a request for computational assistance at any time instant, where the arrival of such requests follows a general stochastic process. For efficient collaboration, an agent deployed at the host MEC server collects system information from all MEC servers to make appropriate decisions on task migration and resource

4

utilization. In this work, we develop a transformer-based PPO algorithm to efficiently learn the optimal collaboration strategy under dynamic system conditions. The MDP model considered in this work is defined as follows. State space: At each decision epoch t, othe system state is n defined as s(t) = X (t), Q(t), H(t), C(t) , where X (t) denotes the task-related information, including task dependencies, the number of subtasks dk,i , total required resource cycles αk,i , and deadlines’ information γk,i , βk,i , Q(t) denotes the queue states of all computing resources of all MEC servers and server helper reputation (i.e., computed by the successful tasks per total assigned tasks), H(t) represents the communication conditions, and C(t) denotes the computation capacities of MEC servers. To capture temporal correlations, we construct an stacking the past L observations s̃(t) = n augmented state by o s(t − L + 1), . . . , s(t) . This sequence is used as the input to the transformer encoder. Transformer Architecture: The proposed agent employs a time-series Transformer encoder to capture temporal dependencies in the evolution of the MEC system. At each decision epoch, the input to the agent is represented as a sequence s̃(t) ∈ RL×F , where F denotes the feature dimension of each observation. Each observation vector is first mapped to a 64dimensional latent representation through a linear projection layer. A learnable positional embedding is then added to preserve the temporal order of the observations. The resulting sequence is processed by a Transformer encoder comprising one encoder layer with four self-attention heads, an embedding dimension of 64, and a feed-forward dimension of 128. The encoder employs GELU activation, pre-layer normalization, and no dropout. The contextualized representations generated for all historical observations are aggregated using mean pooling along the temporal dimension, followed by a layer-normalization operation. The resulting global context vector is shared by the actor and critic. Specifically, the actor consists of a linear policy head that produces logits over the candidate supporter MEC servers, whereas the critic consists of a linear value head that estimates the scalar state value. Therefore, the dimension of the action space is equal to the number of available supporter MEC servers. During action selection, a categorical distribution is constructed directly from the policy logits, and the selected action determines the MEC server to which the task is assigned. The agent is trained using the PPO clipped surrogate objective with generalized advantage estimation. The Adam optimizer is adopted with a learning rate of 10−4 , while the discount factor, GAE parameter, and PPO clipping ratio are set to 0.99, 0.95, and 0.2, respectively. The value-loss coefficient and entropyregularization coefficient are set to 0.5 and 0.005, respectively, and the gradient norm is clipped to a maximum value of 0.5 to improve training stability. Before each policy update, the estimated advantages are normalized and clipped to the range [−10, 10], and the collected transitions are optimized over three PPO epochs. For a fair comparison, the conventional PPO baseline uses the same action space and PPO optimization parameters, but replaces the Transformer encoder with an MLP

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

Algorithm 1 Transformer-Enhanced PPO for MEC Collaboration 1: Initialize the MEC environment and the Transformerenhanced PPO agent. 2: for each decision epoch t do 3: Observe the historical state sequence s̃(t) = {s(t − L + 1), . . . , s(t)}. 4: Feed s̃(t) into the Transformer encoder to obtain the contextual system representation. 5: Use the policy network to select action a(t) ∈ S, corresponding to the selected MEC server. 6: Execute a(t) and observe reward r(t) and next state s(t + 1). 7: Store the current transition for policy updating. 8: if the policy-update condition is satisfied then 9: Estimate the advantages using the value network. 10: Update the policy and value networks using the PPO clipped objective. 11: Clear the stored transitions. 12: end if 13: end for 14: return the trained Transformer-enhanced PPO agent.

that processes the flattened observation sequence. Action space: The action at time t is defined as a(t) ∈ S, where a(t) = s indicates that the selected MEC server s is chosen to process the task. Therefore, the action corresponds to the task migration decision. Reward: In this work, we directly optimize the objective function using deep reinforcement learning (DRL). Therefore, the reward function is carefully designed to be consistent with the objective: P P (j) ω1 i,k φk,i (t) − ω2 i,k,j γk,i (t) P r(t) = , (10) i,k (dk,i + Γk,i ) here ω1 and ω2 are the trade-off weights. State transition: Given the current state s(t) and action a(t), the system evolves to the next state s(t + 1) according to the stochastic task arrival process, user mobility, wireless channel variation, and queue dynamics.
The transition probability is given by Pr s(t + 1) | s(t), a(t) . Lemma 1. The formulated decision process satisfies the Markov property and can therefore be modeled as an MDP. Proof. Let s(t) and a(t) denote the system state and action at decision epoch t, respectively. To prove the Markov property, it is sufficient to show that the conditional distribution of the next state depends only on the current state and action, i.e.,

Pr s(t + 1) | Ht = Pr s(t + 1) | s(t), a(t) , (11) where Ht = {s(0), a(0), . . . , s(t), a(t)} denotes the system history up to time t. By construction, the state s(t) contains all necessary information for future evolution, including task status, dependency conditions, queue states, communication

5

TABLE I. Simulation Parameters Parameter

Value

Cellular bandwidth Number of resource blocks N Bandwidth per RB Bn User transmit power Pu Path loss exponent p Noise power spectral density, N0

10 MHz 10 1 MHz 199.526 mW 3 −174 dBm/Hz

conditions, and available computation resources. The evolution of the system from time t to t + 1 is governed by: i) the stochastic task arrival process, ii) the Markovian user mobility model, iii) wireless channel variations, and iv) queue dynamics under the selected action a(t). All these dynamics depend only on the current state s(t) and action a(t), and are independent of past states and actions given s(t). Therefore, the transition probability satisfies (11), which verifies the Markov property. Hence, the decision process can be modeled as an MDP. Fig. 2 illustrates the overall framework of the proposed transformer-enhanced PPO for MEC collaboration. At each decision epoch, the agent observes a sequence of past system states, denoted as {s(t − L + 1), . . . , s(t)}, which captures the temporal evolution of queue dynamics, task dependencies, and wireless conditions. This historical state sequence is fed into a transformer encoder to extract context-aware representations by modeling both temporal dependencies and cross-server correlations. The output of the transformer is then utilized by the policy and value networks in the PPO framework. Specifically, the policy network (actor) scores each MEC server and selects the most suitable helper server for task migration, while the value network (critic) estimates the long-term expected reward based on the current system context. The selected action is executed in the collaborative MEC environment, where the host MEC offloads tasks that can migrate to candidate helper MEC servers. After task execution, the environment returns the reward and the updated system state, which are used to train the critic and compute the advantage function for policy updates. This closed-loop interaction enables the agent to learn an efficient collaboration strategy by jointly considering system dynamics, resource availability, and task requirements. By leveraging the transformer, the proposed method effectively captures long-term temporal patterns and inter-server relationships, leading to more informed and robust decision-making compared to conventional PPO approaches. IV. N UMERICAL S TUDY A. System Parameters and Baselines The simulated MEC system consists of six servers, each equipped with a heterogeneous number of GPU cards randomly selected from {1, 2, 4, 8, 12}. The computational capability of each GPU card is set to 3.74 × 1012 FLOPS, which is representative of current high-performance accelerators. The network

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

6

TABLE II. Subtasks’ completion rate (Comp.) and remaining adjustment (Rem.) rates across operations and methods. Op.

feed forward Comp.

layer norm Rem.

Comp.

residual add Rem.

Comp.

Rem.

Trans. Trad. No-Migr. Trans. Trad. No-Migr. Trans. Trad. No-Migr. Trans. Trad. No-Migr. Trans. Trad. No-Migr. Trans. Trad. No-Migr. Trans. Trad. No-Migr. Trans. Trad. No-Migr. 0.91 0.64 0.39 0.31 0.26

0.90 0.63 0.38 0.31 0.26

0.90 0.55 0.36 0.23 0.25

0.91 0.70 0.59 0.51 0.46

0.91 0.71 0.60 0.49 0.45

0.89 0.58 0.49 0.38 0.35

0.97 0.88 0.70 0.55 0.49

0.97 0.87 0.68 0.54 0.49

0.97 0.83 0.67 0.46 0.45

0.91 0.70 0.54 0.44 0.40

0.91 0.70 0.56 0.42 0.39

0.89 0.59 0.48 0.33 0.33

coverage is modeled as a two-dimensional square area of size 1000×1000 [m2 ]. User devices are uniformly distributed within this region, with the number of users ranging from 10 to 50 per server on average. Each user generates inference requests associated with LLM workloads following a GPT-3-like architecture. Each subtask corresponds to a component in the model, including layer normalization, multi-head self-attention, feed-forward networks, and residual addition. The prompt length is uniformly distributed between 30 and 50 tokens, while the corresponding transmission data size ranges from 100 to 1000 [KB]. The computational cost of each transformer component varies significantly: multi-head self-attention (with 96 heads) requires approximately 3.2 × 108 to 1.2 × 109 cycles, followed by feed-forward networks with 2.0 × 108 to 3.3 × 108 cycles. In contrast, layer normalization and residual addition incur much lower costs, on the order of 104 to 105 cycles. These values are derived based on standard FLOPs analysis of multi-head selfattention and feed-forward networks [14], [15], and reflect practical implementation variations across different input sizes and system configurations. Additional system and communication parameters are summarized in Table I. These include wireless transmission settings, channel characteristics, and infrastructure configurations. B. Performance Comparison In order to demonstrate the performance of the proposed DRL approach with the system, we design a traditional PPO agent without a transformer encoder component (Traditional PPO). Furthermore, the proposed DRL method is also compared with the system where the migration is not applied (No migration). 1) Convergence of Transformer-enhanced PPO As shown in Fig. 3a, the convergence of the proposed DRL is slower in the first 4000 epochs compared to the traditional PPO, which converges in less than 2000 epochs. However, the average reward curve of the proposed DRL is consistent and exhibits less fluctuation than the baseline scheme. At the end of the training epochs, the proposed DRL’s convergence shows a significant gap with the others. The slower convergence of the proposed method is mainly due to the increased model complexity and the use of historical state sequences, which enlarges the learning space. However, the transformer enables the agent to capture temporal dependencies and helps reduce

1.00 0.98 0.96 0.97 0.95

0.99 0.97 0.92 0.97 0.95

0.99 0.98 0.96 0.97 0.95

0.96 0.85 0.77 0.76 0.75

0.96 0.85 0.79 0.76 0.75

0.95 0.81 0.76 0.72 0.71

0.91 0.66 0.42 0.33 0.28

0.91 0.65 0.41 0.32 0.28

0.90 0.57 0.39 0.25 0.26

0.91 0.72 0.60 0.52 0.46

0.91 0.72 0.62 0.5 0.46

0.89 0.59 0.51 0.40 0.36

1.0 0.9 Average Reward

10 20 30 40 50

Comp.

0.8 0.7 0.6 Transformer-enhanced PPO Traditional PPO

0.5 0.4

0

2000 4000 6000 8000 10000 12000 14000 Epochs

(a) Training performance of the proposed method. 1.0 0.90

0.93

0.92

Traditional PPO Transformer-enhanced PPO

0.89

0.8

rate

Users

self attention Rem.

0.59

0.6

0.60 0.48

0.4

0.37

0.35 0.28

0.2

0.0

10

20

30 Number of Users

40

50

(b) Server completion tasks’ rate versus users.

Fig. 3. Performance comparison of training and task completion. variance through its attention mechanisms, resulting in more stable training behavior and improved long-term performance. In another aspect, we compare the completion rate of migrated tasks in Fig. 3b, the result shows that the proposed transformerenhanced PPO has consistent and significant gaps compared to the other two algorithms. This shows the credibility of the proposed algorithm, identical to the results shown in the learning curve. 2) The completion rate of migration tasks Fig. 4a shows the completion rates of the different methods, which directly impact overall task completion performance. Under different operators, two key characteristics are observed. First, the migration policy increases both the number of subtasks completed during computation and the number of remaining deadline adjustment opportunities in the system. Second, the Transformer-enhanced PPO improves the total number of completed subtasks by 2%, which corresponds to the 9% increase in migrated tasks shown in Fig. 4b and Fig. 3b. Note that these results are obtained with 40 users, where even a

ACCEPTED FOR PRESENTATION AT 2026 IEEE GLOBECOM SELECTED AREAS IN COMMUNICATIONS: CLOUD/EDGE COMPUTING AND NETWORKING

0.97 0.97 0.97

1.0

Completion Rate

0.8

No migration Traditional PPO Transformer-enhanced PPO

0.6

0.54 0.55

rent framework assumes profiled computational workloads and a high-speed inter-MEC backhaul. Future work will consider runtime variability in LLM inference, bursty traffic, backhaul congestion, and extension-duration-aware objectives.

0.46

0.4

R EFERENCES

0.32 0.33

0.31 0.31 0.25

0.23

0.2

0.0

feed_forward

layer_norm residual_add Operator Type

self_attention

(a) Completion rate of subtasks 1.0 Remaining Adjustment Rate

7

0.8

0.72

0.6 0.49

0.4

0.76 0.76

No migration Traditional PPO Transformer-enhanced PPO

0.51

0.50

0.52 0.42 0.44

0.40

0.38

0.33

0.2

0.0

feed_forward

layer_norm residual_add Operator Type

self_attention

(b) Remaining adjustment rate.

Fig. 4. Completion rate and remaining adjustment across operator types (40 users). 1% improvement corresponds to saving millions of subtasks. The completion rate and the remaining number of deadline extensions for subtasks under varying numbers of users are reported in Table II. As the number of users increases, the completion rate consistently decreases across most operations, indicating reduced system effectiveness under higher load. In another aspect, simpler operations such as layer normalization remain highly stable, while more complex components like selfattention and feed-forward networks exhibit significant performance degradation. In addition, the transformer-enhanced PPO maintains higher completion rates and remaining adjustments compared to baseline methods, demonstrating better robustness under scaling. V. C ONCLUSION This work has proposed a learning-based collaborative MEC framework for LLM inference under soft-deadline constraints. Transformer-enhanced PPO approach effectively captures temporal system information as well as cross-server interactions, enabling more informed task-migration decisions. The results demonstrate that the proposed method achieves higher task completion rates with fewer deadline extensions compared to conventional approaches. In the future, we will extend the framework by incorporating adaptive user association and more realistic system dynamics, as well as exploring scalable multiagent learning schemes for large-scale MEC deployments. In addition, integrating resource-aware LLM model compression and dynamic workload partitioning remains a promising direction to further improve system efficiency. Furthermore, the cur-

[1] G. Qu, Q. Chen, W. Wei, Z. Lin, X. Chen, and K. Huang, “Mobile edge intelligence for large language models: A contemporary survey,” IEEE Commun. Surveys Tuts., vol. 27, no. 6, pp. 3820–3860, 2025. [2] J. Li, G. Sun, Z. Sun, J. Wang, Y. Liu, R. Zhang, D. Niyato, and S. Mao, “Llm-guided drl for multi-tier leo satellite networks with hybrid fso/rf links,” IEEE J. Sel. Areas Commun., 2025. [3] M. Zhang, X. Shen, J. Cao, Z. Cui, and S. Jiang, “Edgeshard: Efficient llm inference via collaborative edge computing,” IEEE Internet Things J., vol. 12, no. 10, pp. 13 119–13 131, 2024. [4] E. J. Husom, A. Goknil, M. Astekin, L. K. Shar, A. KÃ¥ sen, S. Sen, B. A. Mithassel, and A. Soylu, “Sustainable llm inference for edge ai: Evaluating quantized llms for energy efficiency, output accuracy, and inference latency,” ACM Transactions on Internet of Things, vol. 6, no. 4, pp. 1–35, 2025. [5] P. Mach and Z. Becvar, “Mobile edge computing: A survey on architecture and computation offloading,” IEEE Commun. Surveys Tuts., vol. 19, no. 3, pp. 1628–1656, 2017. [6] Z. Hao, H. Jiang, S. Jiang, J. Ren, and T. Cao, “Hybrid slm and llm for edge-cloud collaborative inference,” in Proc. Workshop Edge Mobile Found. Models, 2024, pp. 36–41. [7] J. Liu, Y. Du, K. Yang, J. Wu, Y. Wang, X. Hu, Z. Wang, Y. Liu, P. Sun, A. Boukerche et al., “Edge-cloud collaborative computing on distributed intelligence and model optimization: A survey,” IEEE Commun. Surveys Tuts., 2026. [8] Y. Wang, C. Yang, S. Lan, L. Zhu, and Y. Zhang, “End-edge-cloud collaborative computing for deep learning: A comprehensive survey,” IEEE Commun. Surveys Tuts., vol. 26, no. 4, pp. 2647–2683, 2024. [9] Z. Yao, Z. Tang, W. Yang, and W. Jia, “Enhancing LLM qos through cloud-edge collaboration: A diffusion-based multi-agent reinforcement learning approach,” IEEE Trans. Serv. Comput., 2025. [10] N. Yang, S. Chen, H. Zhang, and R. Berry, “Beyond the edge: An advanced exploration of reinforcement learning for mobile edge computing, its applications, and future research trajectories,” IEEE Commun. Surveys Tuts., vol. 27, no. 1, pp. 546–594, 2024. [11] P. Hämäläinen, A. Babadi, X. Ma, and J. Lehtinen, “Ppo-cma: Proximal policy optimization with covariance matrix adaptation,” in 2020 IEEE 30th Int. Workshop Mach. Learn. Signal Process. (MLSP). IEEE, 2020, pp. 1–6. [12] J. Adkins, M. Bowling, and A. White, “A method for evaluating hyperparameter sensitivity in reinforcement learning,” Adv. Neural Inf. Process. Syst., vol. 37, pp. 124 820–124 842, 2024. [13] N. H. Nguyen, V.-D. Nguyen, A. T. Nguyen, N. Van Thieu, H. N. Nguyen, and S. Chatzinotas, “Deadline-aware joint task scheduling and offloading in mobile-edge computing systems,” IEEE Internet Things J., vol. 11, no. 20, pp. 33 282–33 295, 2024. [14] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Łukasz Kaiser, and I. Polosukhin, “Attention is all you need,” in Adv. Neural Inf. Process. Syst. (NeurIPS), 2017, pp. 5998–6008. [15] M. Shoeybi, M. Patwary, R. Puri, P. LeGresley, J. Casper, and B. Catanzaro, “Megatron-lm: Training multi-billion parameter language models using model parallelism,” arXiv preprint arXiv:1909.08053, 2019.

Related documents

• Assessment of Lasso and Ridge models for soil swelling potential prediction.#771
• BigMoMo: Efficient Inference of Large-Scale MoE with Speculative Decoding on Mobile Devices#1084578
• BigMoMo: Efficient Inference of Large-Scale MoE with Speculative Decoding on Mobile Devices#919358
• EZPro-Multi: Contrastive Learning-Enhanced Multi-property Prediction for Enzyme Engineering#229660
• nvidia/OpenMathReasoning#786575