Adaptive AI Agent Placement and Migration in Edge Intelligence Systems

Jianxiong Guo; Jiayi He; Jiong Lou; Liping Qian; Tian Wang; Weijia Jia; Xingdan Wang; Zhiqing Tang

REVIEW 4 major objections 4 minor 1 cited by

Adaptive AI Agent Placement and Migration in Edge Intelligence Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AntLLM cuts edge AI agent deployment latency by ~9.5% and migration cost by ~11.5% by combining ant-colony search with an LLM refinement step.

desk verdict The lightweight-migration idea is sensible, but the paper's key novelty—the LLM refinement step—is unvalidated and the experiments are too thin to support the claimed gains. read the letter →

arxiv 2508.03345 v1 pith:CVRES6QS submitted 2025-08-05 cs.AI

Xingdan Wang , Jiayi He , Zhiqing Tang , Jianxiong Guo , Jiong Lou , Liping Qian , Tian Wang , Weijia Jia This is my paper

classification cs.AI

keywords AIagentplacementedgeintelligencemigrationantcolonyoptimizationLLM-basedagentsqualityofserviceresourceallocationdistributedsystems

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks how LLM-based AI agents, which must plan, call tools, and remember state, should be placed on edge servers and moved as users travel. It claims that a hybrid algorithm, AntLLM, solves both the placement and migration problems better than greedy, polling, or random assignment. The algorithm first uses ant colony optimization to construct low-cost deployment and migration plans, then hands the best plan to an LLM for refinement and validation. On a distributed testbed spanning four edge locations, the paper reports an average reduction of about 10.3-10.6% in total task delay and 38.6-49.6% in resource consumption compared with baselines, with the abstract citing 9.5% lower deployment latency and 11.5% lower migration cost. The contribution is a systematic cost/latency model for agent placement and migration plus an algorithm that exploits the lightweight nature of agent migration, which transfers only memory and configuration rather than a full codebase.

What carries the argument

The load-bearing mechanism is the AntLLM algorithm, which pairs ant colony optimization with a final LLM-based post-optimization step. Each ant constructs a candidate placement or migration plan by probabilistically choosing servers according to a pheromone matrix and a heuristic built from the EAD/EAM cost models; the best plan is then passed to an LLM, marked in the pseudocode as 'Use MLLM to refine and validate S*' (or M*), before becoming the final decision. For migration, the algorithm uses a net-gain objective: expected latency improvement minus migration time and overhead minus dependency damage, and only triggers migration when the user has moved beyond a hop threshold or the hosting server's resources fall below a safety level. Agent migration is kept lightweight by transferring memory and configuration rather than the full agent codebase.

What would settle it

Run the exact same CBIR workloads on the same four edge servers with the LLM post-optimization step removed, leaving only the ant-colony search; if the delay and resource-consumption improvements over Greedy, Polling, and Random are unchanged, the LLM refinement is not responsible for the reported gains and the distinguishing LLM part of AntLLM is unsupported. Conversely, if deleting the LLM step erases the margin, the LLM is doing the work and its prompt and validation logic would need to be reported for the result to be reproducible.

Watch

Extended reading notes

Core claim

The central claim is that the Edge Agent Deployment (EAD) and Edge Agent Migration (EAM) problems can be formulated as resource-constrained optimization problems and solved by an ant-colony heuristic whose best solution is then refined by an LLM. The paper defines total task latency as transmission plus migration plus agent-initiation plus processing latency, and total cost as computing plus storage plus communication cost, and minimizes these subject to memory, storage, and communication capacity constraints. The migration decision compares the latency benefit of moving an agent closer to the user against migration time, overhead cost, and damage to inter-agent dependency latencies. The paper's experiments, run with content-based image retrieval tasks on edge servers distributed across four regions, show AntLLM outperforming Greedy, Polling, and Random baselines on delay and resource metrics, and the authors conclude that their system autonomously places and migrates agents to maintain quality of service in dynamic edge environments.

Load-bearing premise

The load-bearing premise is that the LLM post-optimization step actually improves the ant-colony solution; the paper gives no prompt, no output format, no validation logic, and no ablation that isolates this step, so if the LLM call is arbitrary or unreliable the reported 10.31-10.64% delay and 38.56-49.61% resource gains could come from the ant-colony algorithm alone.

Editorial extensions

If this is right

If the reported results hold, LLM-based agents can be kept close to mobile users at roughly 10% lower total task delay than greedy, polling, or random placement, without sacrificing quality of service.
The 38.6-49.6% resource-consumption reductions imply that the same edge infrastructure can host substantially more agents, or the same agents on cheaper servers, than baseline deployment policies.
The paper's cost/latency model gives later work a concrete objective for joint placement and migration of agents that carry memory and tool dependencies.
The migration policy, which triggers on user movement or resource bottlenecks and migrates only memory and configuration, offers a directly implementable rule for sustaining agent state continuity in edge networks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An obvious next test the paper does not run is an ablation that removes the LLM post-optimization step; without it, a reader cannot tell whether the reported gains come from the ant-colony search, the LLM refinement, or the interaction.
The same post-optimization pattern could transfer to other combinatorial resource-allocation problems in edge computing, such as service caching or container placement, but the transfer is speculative until the LLM step is specified and benchmarked.
Because the LLM refinement presumably incurs its own latency and API cost, a fair comparison would need to include the overhead of calling the LLM inside the decision loop; the current measurements appear to report only the resulting deployment and migration outcomes.
The lightweight migration assumption rests on memory and configuration being the only agent state; agents with large vector stores, tool caches, or local model adapters would need a richer state model before the same gains could be expected.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the paper has a credible problem statement and one genuinely interesting design choice—treating an AI agent's state as memory plus configuration, so migration moves a few MBs, not the whole runtime. That's the right way to think about agent mobility at the edge. The ACO-based placement and migration framework is also standard but cleanly formulated. If the only claim were 'ACO works reasonably for agent placement in a small testbed,' I'd be fine with it.

The trouble is the claim is about AntLLM, the hybrid with LLM post-optimization. The paper never specifies what the LLM does in Algorithms 1 and 2 ('Use MLLM to refine and validate S*'), no prompt, no output schema, no validation logic, and no ablation removing it. So the 10.31-10.64% delay reduction and 38.56-49.61% resource savings cannot be attributed to the hybrid. The LLM could be doing nothing or could be adding noise. This is the load-bearing flaw.

Compounding: the experiments have no error bars or significance tests, the baselines are only Random, Greedy, and Polling, the testbed is 2-4 servers and 5-7 tasks, and the abstract's 9.5%/11.5% improvements don't match the body's numbers. Those last points are fixable in a revision, but they make the current manuscript a hard reject as-is.

What's genuinely good: the system implementation on AgentScope is real engineering, the resource and latency model is sensible, and the idea that agent migration is cheaper than container migration because only memory/config need to move is worth stating explicitly. The paper also cites the relevant prior work on container and service migration, though the references are heavily self-citations—not a problem in itself, but worth a glance.

Who is this for: people working on edge AI orchestration. It's a reasonable system prototype, but not yet a reliable algorithmic contribution. The LLM step needs to be either specified and ablated or dropped. If the authors add a rigorous evaluation with repeated trials, confidence intervals, and a proper ablation, the core idea could be salvageable.

Recommendation: send it to peer review—the topic is timely and the direction is not wrong—but tell the authors clearly that the current evidence doesn't support the central claim. A serious refereeing process could force the missing ablation and statistical rigor. Without that, I wouldn't rely on any of the numbers.

Referee Report

4 major / 4 minor

Summary. The paper formulates the Edge Agent Deployment (EAD) and Edge Agent Migration (EAM) problems for LLM-based AI agents in edge intelligence systems, and proposes AntLLM, a hybrid algorithm that combines ant colony optimization with an LLM-based post-optimization step. The authors implement the system using AgentScope on Huawei Cloud servers in four regions, evaluate it on CBIR tasks, and compare it with Random, Greedy, and Polling baselines. They report average total-delay reductions of 10.31% (varying the number of servers) and 10.64% (varying the number of tasks), and resource-consumption reductions of 38.56% and 49.61%.

Significance. If the results are reproducible, the paper would offer a practical and timely contribution to edge AI agent lifecycle management, with a useful emphasis on agent-specific state transfer and resource-aware placement. The real distributed implementation with AgentScope is a strength, as is the explicit treatment of memory/configuration migration rather than full container migration. However, the central quantitative claims are currently not well supported: the LLM refinement step is unspecified and unablated, the stochastic experiments lack repeated trials and error bars, and the baseline set is weak. The significance therefore remains conditional on a substantial revision that pins down and validates the LLM component and the statistical evidence.

major comments (4)

[§III-A, Algorithm 1 line 10; §III-B, Algorithm 2 line 13] The only element that distinguishes AntLLM from a standard ant colony optimization is the post-optimization step reading 'Use MLLM to refine and validate S*/M*'. The manuscript gives no prompt template, input/output schema, validation rules, failure behavior, or cost model for the LLM calls, and the parameters of the LLM refinement (e.g., number of samples, temperature, validation criteria) are not listed. No ablation removes the LLM step, so the reported gains in Section V cannot be attributed to the AntLLM hybrid rather than to ACO alone or to run-to-run variation. A revision must specify the LLM protocol and include an ablation comparing ACO with and without the LLM refinement.
[§V-B, Figs. 2–5] All delay and cost comparisons are reported as point averages with no error bars, confidence intervals, or significance tests, even though both AntLLM and Random are randomized procedures. The claim 'reduces the total delay by an average of 10.31%' (Section V-B) is only meaningful if the spread across repeated trials is reported; with the current data the improvement is not shown to exceed noise. The authors should provide repeated runs with distinct seeds, per-point variance, and a paired significance test or effect-size measure.
[§V-A] The baseline set contains only Random, Greedy, and Polling. For the small instance sizes used (2–4 servers and 5–7 tasks), an optimal or near-optimal reference (e.g., ILP or exhaustive search for the static EAD problem) is feasible and would calibrate how far the heuristic is from optimality, giving the 'significantly reduces' claim a meaningful absolute reference. Without such a baseline, the reported 10.31%/10.64% improvements are only relative to very weak competitors and do not establish that AntLLM is a strong solver.
[Abstract and §V-B] The abstract reports average reductions of 9.5% deployment latency and 11.5% migration cost, but Section V-B reports total-delay reductions of 10.31% and 10.64% and resource-consumption reductions of 38.56% and 49.61%. No experiment or table in Section V reports the 9.5%/11.5% figures, and the relationship between 'deployment latency', 'migration cost', and the metrics in Figs. 2–5 is not defined. The claims should be reconciled and tied to specific figures, or the abstract should be corrected.

minor comments (4)

[Figs. 2–5] Several axis labels are garbled, e.g., 'G eedy', 'Greed', 'T a k', 'Number f N des', 'P lling', 'Rand m', 'A tLLM', 'Memo y Usage', 'VM U age', and 'VMs Usage'. Please regenerate the figures with clean labels.
[§III-B] The text says 'The ALM algorithm is illustreted in Algorithm 2'; this should be 'illustrated'. Also, the algorithm boxes use 'MLLM' while the abstract and introduction use 'LLM'; please unify the terminology.
[§II-B, Eqs. (7)–(10)] The cost components in Eqs. (7)–(10) sum quantities with different units (seconds, MB, Mbps) into a single total cost O_r. The objective and constraint formulation would benefit from explicit normalization or weights, and the displayed statement of Problem EAD appears incomplete, with an unclosed summation in the objective. Please provide the full, well-formed optimization problem.
[§V-A] The manuscript does not list the numerical values of key parameters (alpha, beta, rho, Q, number of ants, number of iterations, H_th, gamma, theta_dep, and the LLM configuration), nor does it state whether code or data will be released. Without these, the experiments are not reproducible even if the algorithmic description were complete.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the objectives, algorithm, and evaluation metrics are independently defined; the unablated LLM refinement step is a reproducibility concern, not a circular reduction.

full rationale

The paper's derivation chain is self-contained with respect to circularity. The EAD/EAM objective functions (Problems EAD/EAM) and the total delay/cost terms (Eqs. (1)-(10)) are defined directly from measurable latency, migration, storage, and communication quantities, and the ACO selection rules (Eqs. (12), (14)), pheromone updates (Eqs. (13), (16), (17)), and fitness functions (Eq. (15)) all minimize these same independently defined objectives. No parameter is fitted to the evaluation data and then renamed as a prediction; the reported delay reductions and resource-usage reductions are comparisons against Greedy, Polling, and Random baselines on the same metrics. Prior work by the authors is cited only as background on container migration and edge computing, not as an authority that forces the algorithm choice, and no uniqueness theorem is imported from a self-citation. The LLM post-optimization step (Algorithm 1 line 10 and Algorithm 2 line 13, 'Use MLLM to refine and validate S*/M*') is underspecified and never ablated, so the marginal contribution of the LLM component is not established; that is a correctness, reproducibility, and attribution gap, not circularity, because no equation or fitted value makes the reported result equal to its input by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper relies on standard ant colony optimization plus an under-specified LLM post-processing step. No new physical or conceptual entities are introduced. The main burden is on the unspecified LLM behavior and the unreported algorithmic parameters, which prevent independent verification of the central claim.

free parameters (4)

ACO parameters (alpha, beta, rho, Q, number of ants, iterations) = not reported
The ant colony algorithm's performance depends on these values; without them the reported improvements cannot be independently reproduced or assessed for sensitivity.
EAM net gain weights gamma and theta_dep = not reported
Used in Problem EAM to balance latency benefit, migration cost, and dependency damage; the values are not specified.
Migration trigger threshold H_th and resource bottleneck threshold = not reported
These conditions trigger the EAM migration decision and are not quantified.
LLM prompt and validation configuration = not reported
The LLM refinement step is central to the AntLLM name but no prompt, model parameters, or validation procedure are disclosed.

assumptions (5)

domain assumption Edge server resource capacities are known and static during a decision epoch.
The placement and migration decisions assume accurate real-time resource reads via Paramiko and unchanged capacities during decision execution.
domain assumption Network bandwidth between servers and to users is stable and known.
Used in latency equations (1) and (2); no dynamic bandwidth model or variance is considered.
domain assumption AI agent resource requirements (memory, computing, storage, communication) are fixed and known.
Task analysis generates resource requirements; no uncertainty or workload variability is modeled.
domain assumption Migration of an AI agent requires transferring only memory and configuration, not the codebase.
Core assumption enabling lightweight migration; not empirically validated for correctness or cold-start effects.
ad hoc to paper The LLM used in post-optimization yields valid and beneficial placement or migration adjustments.
The LLM refinement step is invoked but its mechanism, prompt, and validation criteria are unspecified, making this a load-bearing unverified assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive AI Agent Placement and Migration in Edge Intelligence Systems." pith.science (2026). https://pith.science/paper/CVRES6QS

@misc{pith2026250803345,
  author       = {Pith},
  title        = {Pith review of: Adaptive AI Agent Placement and Migration in Edge Intelligence Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CVRES6QS}},
  note         = {Machine review of arXiv:2508.03345}
}

read the original abstract

The rise of LLMs such as ChatGPT and Claude fuels the need for AI agents capable of real-time task handling. However, migrating data-intensive, multi-modal edge workloads to cloud data centers, traditionally used for agent deployment, introduces significant latency. Deploying AI agents at the edge improves efficiency and reduces latency. However, edge environments present challenges due to limited and heterogeneous resources. Maintaining QoS for mobile users necessitates agent migration, which is complicated by the complexity of AI agents coordinating LLMs, task planning, memory, and external tools. This paper presents the first systematic deployment and management solution for LLM-based AI agents in dynamic edge environments. We propose a novel adaptive framework for AI agent placement and migration in edge intelligence systems. Our approach models resource constraints and latency/cost, leveraging ant colony algorithms and LLM-based optimization for efficient decision-making. It autonomously places agents to optimize resource utilization and QoS and enables lightweight agent migration by transferring only essential state. Implemented on a distributed system using AgentScope and validated across globally distributed edge servers, our solution significantly reduces deployment latency and migration costs.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Cost and Accuracy of Long-Term Memory in Distributed Multi-Agent Systems Based on Large Language Models
cs.IR 2026-01 reject novelty 5.0 of 10

A two-framework testbed comparison claims mem0 is Pareto-optimal over Graphiti for distributed LLM agents because its lower cost is paired with accuracy that is not significantly different.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages · cited by 1 Pith paper

[1]

Energy-efficient joint task assignment and migration in data centers: A deep reinforcement learning approach,

J. Lou, Z. Tang, and W. Jia, “Energy-efficient joint task assignment and migration in data centers: A deep reinforcement learning approach,” IEEE Transactions on Network and Service Management, vol. 20, no. 2, pp. 961–973, 2022

work page 2022
[2]

Adaptive digital twin migration in vehicular edge computing and networks,

F. Mou, J. Lou, Z. Tang, Y . Wu, W. Jia, Y . Zhang, and W. Zhao, “Adaptive digital twin migration in vehicular edge computing and networks,” IEEE Transactions on Vehicular Technology, 2024

work page 2024
[3]

Multi-user layer- aware online container migration in edge-assisted vehicular networks,

Z. Tang, F. Mou, J. Lou, W. Jia, Y . Wu, and W. Zhao, “Multi-user layer- aware online container migration in edge-assisted vehicular networks,” IEEE/ACM Transactions on Networking, vol. 32, no. 2, pp. 1807–1822, 2023

2023
[4]

Migration mod- eling and learning algorithms for containers in fog computing,

Z. Tang, X. Zhou, F. Zhang, W. Jia, and W. Zhao, “Migration mod- eling and learning algorithms for containers in fog computing,” IEEE Transactions on Services Computing, vol. 12, no. 5, pp. 712–725, 2018

work page 2018
[5]

Agentscope: A flexible yet robust multi- agent platform,

D. Gao, Z. Li, X. Pan, W. Kuang, Z. Ma, B. Qian, F. Wei, W. Zhang, Y . Xie, D. Chen, L. Yao, H. Peng, Z. Zhang, L. Zhu, C. Cheng, H. Shi, Y . Li, B. Ding, and J. Zhou, “Agentscope: A flexible yet robust multi- agent platform,” 2024

work page 2024
[6]

Edge intelligence: Paving the last mile of artificial intelligence with edge computing,

Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelligence: Paving the last mile of artificial intelligence with edge computing,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1738–1762, 2019

work page 2019
[7]

Collaborative service placement for edge computing in dense small cell networks,

L. Chen, C. Shen, P. Zhou, and J. Xu, “Collaborative service placement for edge computing in dense small cell networks,” IEEE Transactions on Mobile Computing, vol. 20, no. 2, pp. 377–390, 2021

work page 2021
[8]

Deepseek-v3 technical report,

DeepSeek-AI and et al., “Deepseek-v3 technical report,” 2025

work page 2025

Show all 17 references

[9]

Latency minimization for intelligent reflecting surface aided mobile edge computing,

T. Bai, C. Pan, Y . Deng, M. Elkashlan, A. Nallanathan, and L. Hanzo, “Latency minimization for intelligent reflecting surface aided mobile edge computing,” IEEE Journal on Selected Areas in Communications, vol. 38, no. 11, pp. 2666–2682, 2020

2020
[10]

Enhancing LLM QoS through Cloud-Edge Collaboration: A Diffusion-based Multi-Agent Reinforce- ment Learning Approach ,

Z. Yao, Z. Tang, W. Yang, and W. Jia, “ Enhancing LLM QoS through Cloud-Edge Collaboration: A Diffusion-based Multi-Agent Reinforce- ment Learning Approach ,” IEEE Transactions on Services Computing, no. 01, pp. 1–17, Apr. 2025, doi: 10.1109/TSC.2025.3562362

2025
[11]

Edsp-edge: Efficient dynamic edge service entity placement for mobile virtual reality systems,

X. Chi, H. Chen, G. Li, Z. Ni, N. Jiang, and F. Xia, “Edsp-edge: Efficient dynamic edge service entity placement for mobile virtual reality systems,” IEEE Transactions on Wireless Communications, vol. 23, no. 4, pp. 2771–2783, 2024

2024
[12]

A survey on edge computing systems and tools,

F. Liu, G. Tang, Y . Li, Z. Cai, X. Zhang, and T. Zhou, “A survey on edge computing systems and tools,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1537–1562, 2019

2019
[13]

Multiagent deep rein- forcement learning for cost- and delay-sensitive virtual network function placement and routing,

S. Wang, C. Yuen, W. Ni, Y . L. Guan, and T. Lv, “Multiagent deep rein- forcement learning for cost- and delay-sensitive virtual network function placement and routing,” IEEE Transactions on Communications, vol. 70, no. 8, pp. 5208–5224, 2022

2022
[14]

Greengard, The internet of things

S. Greengard, The internet of things. MIT press, 2021

2021
[15]

Edge- based differential privacy computing for sensor–cloud systems,

T. Wang, Y . Mei, W. Jia, X. Zheng, G. Wang, and M. Xie, “Edge- based differential privacy computing for sensor–cloud systems,” Journal of Parallel and Distributed Computing, vol. 136, pp. 75–85, 2020

2020
[16]

Edge computing for autonomous driving: Opportunities and challenges,

S. Liu, L. Liu, J. Tang, B. Yu, Y . Wang, and W. Shi, “Edge computing for autonomous driving: Opportunities and challenges,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1697–1716, 2019

2019
[17]

Ai augmented edge and fog computing: Trends and challenges,

S. Tuli, F. Mirhakimi, S. Pallewatta, S. Zawad, G. Casale, B. Javadi, F. Yan, R. Buyya, and N. R. Jennings, “Ai augmented edge and fog computing: Trends and challenges,” Journal of Network and Computer Applications, vol. 216, p. 103648, 2023

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Energy-efficient joint task assignment and migration in data centers: A deep reinforcement learning approach,

J. Lou, Z. Tang, and W. Jia, “Energy-efficient joint task assignment and migration in data centers: A deep reinforcement learning approach,” IEEE Transactions on Network and Service Management, vol. 20, no. 2, pp. 961–973, 2022

work page 2022

[2] [2]

Adaptive digital twin migration in vehicular edge computing and networks,

F. Mou, J. Lou, Z. Tang, Y . Wu, W. Jia, Y . Zhang, and W. Zhao, “Adaptive digital twin migration in vehicular edge computing and networks,” IEEE Transactions on Vehicular Technology, 2024

work page 2024

[3] [3]

Multi-user layer- aware online container migration in edge-assisted vehicular networks,

Z. Tang, F. Mou, J. Lou, W. Jia, Y . Wu, and W. Zhao, “Multi-user layer- aware online container migration in edge-assisted vehicular networks,” IEEE/ACM Transactions on Networking, vol. 32, no. 2, pp. 1807–1822, 2023

2023

[4] [4]

Migration mod- eling and learning algorithms for containers in fog computing,

Z. Tang, X. Zhou, F. Zhang, W. Jia, and W. Zhao, “Migration mod- eling and learning algorithms for containers in fog computing,” IEEE Transactions on Services Computing, vol. 12, no. 5, pp. 712–725, 2018

work page 2018

[5] [5]

Agentscope: A flexible yet robust multi- agent platform,

D. Gao, Z. Li, X. Pan, W. Kuang, Z. Ma, B. Qian, F. Wei, W. Zhang, Y . Xie, D. Chen, L. Yao, H. Peng, Z. Zhang, L. Zhu, C. Cheng, H. Shi, Y . Li, B. Ding, and J. Zhou, “Agentscope: A flexible yet robust multi- agent platform,” 2024

work page 2024

[6] [6]

Edge intelligence: Paving the last mile of artificial intelligence with edge computing,

Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelligence: Paving the last mile of artificial intelligence with edge computing,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1738–1762, 2019

work page 2019

[7] [7]

Collaborative service placement for edge computing in dense small cell networks,

L. Chen, C. Shen, P. Zhou, and J. Xu, “Collaborative service placement for edge computing in dense small cell networks,” IEEE Transactions on Mobile Computing, vol. 20, no. 2, pp. 377–390, 2021

work page 2021

[8] [8]

Deepseek-v3 technical report,

DeepSeek-AI and et al., “Deepseek-v3 technical report,” 2025

work page 2025