Multi-Turn Reasoning LLMs for Task Offloading in Mobile Edge Computing
Authors: Ning Yang, Chuangxin Cheng, Haijun Zhang
Organizations: Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China · School of Mechano-Electronic Engineering, Xidian University, Xi’an, 710071, China · School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing, 100083, China
Emerging computation-intensive applications impose stringent latency requirements on resource-constrained mobile devices. Mobile Edge Computing (MEC) addresses this challenge through task offloading. However, designing effective policies remains difficult due to dynamic task arrivals, time-varying channels, and the spatio-temporal coupling of server queues. Conventional heuristics lack adaptability, while Deep Reinforcement Learning (DRL) suffers from limited generalization and architectural rigidity, requiring retraining when network topology changes. Although Large Language Models (LLMs) offer semantic reasoning capabilities, standard Supervised Fine-Tuning (SFT) yields myopic policies that greedily minimize immediate latency without accounting for long-term system evolution. To address these limitations, we propose COMLLM, a generative framework that enables foresighted decision-making in MEC systems. COMLLM integrates Group Relative Policy Optimization (GRPO) with a Look-Ahead Collaborative Simulation (LACS) mechanism, which performs multi-step Monte Carlo rollouts while jointly modeling server queue dynamics. By incorporating these rollouts into the reward design, the framework captures the long-term impact of current decisions on future system states. Experimental results demonstrate that COMLLM achieves near-optimal latency and improved load-balancing fairness. Notably, it exhibits zero-shot topological scalability, allowing a model trained on small-scale networks to generalize to larger, unseen topologies without retraining, outperforming SFT, DRL, and heuristic baselines.
Figures & tables
Fig. 1: Overview of COMLLM. MEC states are serialized into textual prompts for SFT initialization. During GRPO training, candidate actions are evaluated using the physical MEC cost and LACS, whereas deployment requires only a single policy inference.
Category
Parameter
Value
System Environment
Time slot duration Δt
0.1 s
Local CPU frequency floc
2 GHz
Edge server CPU frequency Fe
[20.0,48.0] GHz
Average uplink rate Ru,eup
14 Mbps
Task size Dt
[2.0,5.0] Mbits
Computational density γ
0.297 gigacycles/Mbit
TABLE I: Parameter settings in the experimental setup
Model
Average Latency
Drop Rate
Performance Ratio
Load Balancing Index
(time slots)
(%)
(%)
(%)
COMLLM
3.0745
0.00
96.86
73.87
GRPO-7B
3.1197
0.00
95.46
71.20
DQN
3.3966
4.35
87.68
65.64
SFT-7B
4.0989
0.33
72.65
41.39
GRPO-1.5B
4.3096
1.63
69.10
19.94
TABLE II: Overall performance comparison of offloading policies
Model
Task Size 2 Mbits
Task Size 4 Mbits
Task Size 6 Mbits
Task Size 8 Mbits
Task Size 10 Mbits
Average Latency
Drop Rate
Average Latency
Drop Rate
Average Latency
Drop Rate
Average Latency
Drop Rate
Average Latency
Drop Rate
(time slots)
(%)
(time slots)
(%)
(time slots)
(%)
(time slots)
(%)
(time slots)
(%)
COMLLM
1.8515
0.00
3.7409
0.00
5.3290
0.00
6.9759
1.08
9.5363
2.78
GRPO-7B
1.9723
0.00
3.8420
0.00
5.4595
0.00
7.1910
2.16
9.8527
4.31
SFT-7B
3.1702
0.00
4.9999
0.00
6.9515
3.03
9.2281
31.11
12.4634
55.02
GRPO-1.5B
2.7558
0.00
5.1860
2.00
7.0620
8.08
8.5224
20.00
11.4648
28.23
TABLE III: Performance under different task workload levels
Model
3 Servers
5 Servers
7 Servers
9 Servers
11 Servers
Average Latency
Drop Rate
Perf. Ratio
Average Latency
Drop Rate
Perf. Ratio
Average Latency
Drop Rate
Perf. Ratio
Average Latency
Drop Rate
Perf. Ratio
Average Latency
Drop Rate
Perf. Ratio
(time slots)
(%)
(%)
(time slots)
(%)
(%)
(time slots)
(%)
(%)
(time slots)
(%)
(%)
(time slots)
(%)
(%)
COMLLM
3.4097
0.00
93.90
3.0688
0.00
96.99
2.9026
0.00
98.59
2.8447
0.00
96.64
3.1114
0.00
97.42
GRPO-7B
3.4527
0.00
92.73
3.0718
0.00
96.90
2.9123
0.00
98.26
2.8518
0.00
96.41
3.1174
0.00
97.23
SFT-7B
4.2683
0.00
75.01
4.3316
0.00
68.72
3.5708
0.00
80.14
3.8498
0.00
71.41
4.4130
0.00
68.69
GRPO-1.5B
4.6658
2.54
68.62
4.3707
0.96
68.10
4.3243
1.70
66.18
3.8503
0.00
71.40
4.7216
2.76
64.20
TABLE IV: Performance under different edge-server topologies
Model
Local
Server1
Server2
Server3
Server4
Server5
Server6
Server7
Server8
Server9
Server10
Server11
Load Bal. Index(%)
COMLLM
6
14
52
62
26
27
18
39
19
19
6
18
73.87
GRPO-7B
1
16
53
66
27
26
19
39
22
19
0
18
71.20
DQN
16
18
48
10
64
50
35
25
16
10
10
4
65.64
SFT-7B
136
9
67
30
17
4
11
1
0
3
9
19
41.39
GRPO-1.5B
12
192
43
10
19
10
1
0
4
10
0
5
19.94
Random
59
46
42
51
26
31
16
15
10
6
1
3
63.42
TABLE V: Offloading distribution and load balancing index
Model
Standard Env.
Shuffled Params.
Noisy Env.
Unit Variation
Average Latency
Drop Rate
Load Bal. Index
Average Latency
Drop Rate
Load Bal. Index
Average Latency
Drop Rate
Load Bal. Index
Average Latency
Drop Rate
Load Bal. Index
(time slots)
(%)
(%)
(time slots)
(%)
(%)
(time slots)
(%)
(%)
(time slots)
(%)
(%)
COMLLM
3.0675
0.00
94.71
3.0684
0.00
95.27
3.0617
0.00
94.18
3.0620
0.00
92.67
GRPO-7B
3.0722
0.00
93.20
3.0836
0.00
95.13
3.0868
0.00
93.12
3.0885
0.00
92.14
SFT-7B
4.3581
0.00
85.23
4.2393
0.00
87.76
4.1582
0.00
90.14
4.1999
0.00
84.94
GRPO-1.5B
4.3831
1.91
74.29
4.3666
1.47
69.07
4.3879
1.91
77.43
4.3710
0.48
69.21
TABLE VI: Robustness under semantic prompt perturbations
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 decision-making 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.
Ngoc Hung Nguyen, Bjorn Landfeldt
Department of Electrical and Information Technology, Lund University, P.O. Box 117, Lund 22100, Sweden
With the rapid advancement of artificial intelligence (AI) and intelligent science, intelligent edge computing has been widely adopted. However, the limitations of traditional methods, such as poor adaptability and the slow convergence of heuristic algorithms, are becoming increasingly evident. To enable sustainable and resource-efficient edge applications, this paper proposes an online task offloading framework for wireless powered mobile edge computing (MEC) networks, called Quantum Attention-based Reinforcement learning for Online Offloading (QAROO). The system employs a binary offloading strategy with the aim of co-optimizing computing and energy resources in dynamic channel environments. In response to the issues of poor adaptability in traditional approaches and the slow convergence of heuristic algorithms, the framework integrates quantum neural networks and attention mechanisms, introducing three key improvements: using recurrent neural networks to enhance temporal modeling capability, proposing an uncertainty-guided quantization method to improve exploration efficiency, and incorporating attention mechanisms into quantum networks to strengthen feature representation. Experiments demonstrate that the proposed method outperforms comparative schemes in terms of normalized computation speed and processing time, offering an efficient and stable solution for online task offloading in large-scale Internet of Things (IoT) dynamic environments.
Ensuring strict tail-latency service-level objectives (SLOs) in dynamic mobile edge computing (MEC) systems remains challenging because user mobility, wireless fading, bursty workloads, and partial observability jointly undermine reliable cloud-edge orchestration. Existing microservice migration methods predominantly optimize average delay and often decouple migration from bandwidth control, leading to uncoordinated decisions, queue oscillation, and frequent high-percentile latency violations. To address this issue, we propose IMPACT, an intent-driven Agentic AI framework for cooperative microservice migration and bandwidth control in cloud-edge systems. Under centralized training with decentralized execution (CTDE), each edge cloud is modeled as an autonomous agent that encodes local SLO risk, migration urgency, and computational pressure into compact, semantic intent representations. IMPACT further introduces a double-attention mechanism that first selectively aggregates relevant peer intents for efficient inter-agent communication and then filters local observations to emphasize goal-relevant state information. This design enables robust coordination under partial observability and jointly optimizes service migration and discrete uplink bandwidth allocation. Extensive experiments in 5-edge and 20-edge scenarios show that IMPACT reduces mean latency by 30-50% and tail-latency deviation by 40-70% compared with state-of-the-art factorized multi-agent reinforcement learning (MARL) and heuristic baselines, while achieving near-zero SLO violation rates under tight thresholds and energy consumption close to the best heuristic baseline. These results demonstrate that intent-driven agentic coordination provides an effective and scalable solution for SLO-aware orchestration in complex cloud-edge intelligent systems.
Xinjin Li, Siru Tao, Shihan Yin +6
Columbia University · Carnegie Mellon University · Brandeis University +5