Pith. sign in

REVIEW 4 major objections 5 minor 27 references

For real-time edge AI, this paper argues that a foundation model's layer split and node placement should be decided at runtime, not fixed in advance: an orchestration loop that re-splits and re-places model segments can beat any static conf

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A framework for runtime re-splitting and re-placement of foundation model layers across edge nodes is proposed, but its claimed latency gains are inherited from prior work rather than measured.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A coherent architecture/vision paper whose quantitative claims are inherited from other systems and internally contradicted by its own table — no measurement, no derivation, so not publishable as a research contribution. the 4 major comments →

arxiv 2512.01039 v1 pith:XPLA4N7D submitted 2025-11-30 cs.DC cs.LGcs.NI

Joint Partitioning and Placement of Foundation Models for Real-Time Edge AI

Aladin Djuhera , Fernando Koch , Alecio Binotto This is my paper
classification cs.DC cs.LGcs.NI
keywords foundation modelsedge AIsplit inferencemodel partitioningorchestration6G MEClatencyplacement
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 is trying to establish that static layer partitioning of large foundation models is the wrong default for volatile edge networks, and that a monitor-driven orchestrator which re-splits and re-places model segments at runtime can meet strict latency targets that static splits miss. For a sympathetic reader, the claim is that treating both model segmentation and node placement as runtime-resolved variables yields concrete wins: 39 to 60 percent lower end-to-end latency, up to 2.1 times the throughput, and GPU utilization near 90 percent, all while preserving privacy by keeping sensitive layers on trusted nodes. The central scenario is a 5G/6G MEC deployment serving Llama3-8B across three A100 nodes with backhaul bandwidth varying from 20 to 200 Mb/s. The paper's quantitative case is presented as expected performance, based on an analytical latency model plus improvement margins inherited from other split-inference systems rather than direct measurement of the proposed design.

Core claim

The paper's central claim is that adaptive split inference consistently outperforms any static configuration in heterogeneous, bandwidth-variable edge environments. It formalizes orchestration as a constrained optimization over layer-wise assignments, with a cost function combining latency, resource-usage imbalance, and privacy penalties, subject to unique-assignment, capacity, and privacy constraints. The framework reacts to changing conditions through a monitoring loop: when latency exceeds a threshold, utilization spikes, or bandwidth drops, the orchestrator migrates partitions or invokes a split-revision step to find a new layer segmentation. According to the paper, this keeps end-to-end

What carries the argument

The load-bearing object is the cost function Φ = αL + βU + γP over placement matrix x and split set S, minimized under unique-assignment, capacity, and privacy constraints. The orchestration is driven by a closed feedback loop: Monitoring & Capacity Profiling gathers real-time node and network metrics; the Adaptive Orchestrator decides whether to keep, redistribute, or fully re-split; Split Revision selects a new segmentation S*; and Reconfiguration Broadcast distributes the new assignment. The default trigger thresholds — 150 ms EWMA latency, 85 percent utilization, 50 Mbps minimum bandwidth, and a 30 s cool-down — determine when reconfiguration fires.

Load-bearing premise

The entire performance case depends on the transferability of latency-improvement margins reported by other split-inference systems, because the paper derives its adaptive latency by subtracting those inherited deltas from a static-split base rather than measuring its own monitoring, re-splitting, and reconfiguration design.

What would settle it

Run Llama3-8B on three A100 MEC nodes with a static three-way split and with the proposed adaptive controller at 20, 50, 100, and 200 Mb/s backhaul, and measure end-to-end latency. If the static split ever matches or beats the adaptive scheme, or if the adaptive scheme misses the 150 ms bound at any load, the central claim fails. A second check is whether the inherited improvement deltas actually match the observed difference between static and adaptive latency.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the adaptive scheme performs as modeled, a single monitor-driven loop can keep end-to-end inference under the 150 ms URLLC bound even at 20 Mb/s backhaul, where the static split is estimated at 500 ms.
  • The expected improvements over static splitting are 39 to 60 percent lower latency and up to 2.1 times baseline throughput across the 20–200 Mb/s sweep, with diminishing returns above roughly 100 Mb/s.
  • Privacy can be bundled at no extra orchestration cost by constraining input-near layers to trusted nodes, so sensitive data never leaves a trusted domain.
  • Because decisions act on the model's computational graph rather than on opaque containers, the same substrate would apply across different LFM architectures without model-specific redesign.
  • The small monitored overhead — up to 10 ms per cycle — is claimed to be amortized by hundreds of milliseconds saved per request, making reconfiguration net-positive in practice.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's headline numbers are not measured for its own design: adaptive latency is computed as static-split latency minus deltas inherited from other systems. A direct benchmark of Llama3-8B on three A100 MEC nodes would either confirm or overturn the 150 ms claim.
  • If the approach works, the same feedback loop could extend to multi-tenant MEC scenarios where competing inference workloads, not just backhaul drops, cause utilization spikes and stragglers — a situation the current single-model simulation does not exercise.
  • The 'privacy at no cost' property depends on input-near layers being cheap enough to run on the trusted node; for very large models this may not hold, in which case privacy preservation would trade directly against latency.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes an adaptive split-inference orchestration framework for foundation models in edge/MEC environments. It introduces a reference architecture (Monitoring/Capacity Profiling, Adaptive Orchestrator, Split Revision, Reconfiguration Broadcast), formulates an optimization problem over layer-wise placement and splitting decisions subject to latency, utilization, and privacy constraints, and reports expected performance results (Table II) for a simulated 5G-MEC scenario with Llama3-8B on A100 GPUs. The central claim is that adaptive, runtime re-splitting and re-placement consistently outperforms static partitioning, reducing end-to-end latency by 39–60% and meeting a 150 ms QoS bound under varying backhaul bandwidth.

Significance. If the quantitative claims were established, the paper would address a real and timely problem: runtime reconfiguration of large-model splits in volatile edge networks. The conceptual contribution — treating split boundaries and placement as jointly dynamic decision variables — is reasonable, and the modular architecture (CP/AO/SR/RB) plus the formal statement of constraints (Eqs. 1–9) provide a useful skeleton for future work. However, the paper's load-bearing performance numbers are not derived, measured, or independently reproducible: Table II is generated by combining the ETSI MEC model with improvement margins 'inherited' from other papers, and the constants in the latency model are never specified. The paper also contains an internal contradiction between its 150 ms QoS bullet and the 200 ms adaptive latency reported at 20 Mb/s. Because the central claims rest on unverified and partially inconsistent quantitative support, the significance is currently unrealized.

major comments (4)
  1. [Section IV, Eq. (10) and Table II] The quantitative core is not reproducible. Eq. (10) defines latency as T_proc + T_queue + T_tx(bandwidth), but T_proc and T_queue are never given values or derivations. Table II then reports adaptive latencies (200, 150, 120, 110 ms) that are supposed to follow from this model plus margins 'inherited from related works [24]–[26]' and scaled with the analytical latency model. No equation connects the inherited margins to the proposed CP/AO/SR/RB orchestration, to Llama3-8B, or to an A100-based 3-node MEC topology. The claimed −60% to −39% improvement and the 2.1× throughput figure are therefore unsupported premises rather than results of this paper.
  2. [Section IV, QoS bullet vs. Table II] The bullet 'the 150 ms URLLC bound is met across all loads only under the adaptive scheme' is directly contradicted by Table II, which lists a 200 ms adaptive latency at 20 Mb/s backhaul. Even if 150 ms were meant only for loads at ≥50 Mb/s, the statement as written is false, and this inconsistency undermines the central robustness claim.
  3. [Section III-C, Eq. (6)–(8)] The optimization formulation is not instantiated or analyzed. Eq. (6) defines a min over S∈Ω and x, but the paper provides no algorithm for solving it, no characterization of Ω, no cost model for U or P, and no analysis of the trade-off among α, β, γ. The authors explicitly say 'We leave the investigation of optimal strategies to future work.' The trigger thresholds in Table I (150 ms, 0.85, 50 Mbps, 30 s) are stated without derivation or sensitivity analysis. Thus the workflow in Alg. 1 is a plausible architecture sketch but not a validated decision mechanism.
  4. [Section IV, methodology and transferability] The performance evaluation relies entirely on figures 'reported across four public studies' and on deltas that 'inherit margins from related works [24]–[26]' and '[17]'. These cited systems use different models, hardware, and split strategies; no argument is given for why their improvement margins transfer to this framework. The paper itself concedes that 'comprehensive cross-framework benchmarking' is future work. As a result, the conclusion 'Adaptive split inference thus consistently outperforms any static configuration' is not supported by the evidence presented.
minor comments (5)
  1. [Section III-B, Eq. (3)] The placement matrix notation is inconsistent: the text says rows correspond to nodes in N ∪ {c}, but Eq. (3) sums over i∈N and treats x_c,j separately, without showing the cloud index in the matrix definition. Clarify whether x is (n+1)×k or whether the cloud is handled separately.
  2. [Algorithm 1] The pseudocode is incomplete: 'ShouldReconfigure(E(t),Θ)' is called, but the function is not formally defined; the trigger logic in the text is stated informally. Also, the 'foreach' loop and the 'if' block make it unclear whether a reconfiguration can occur more than once per monitoring window.
  3. [Section IV, heading] Caption/heading says '10S monitoring window'; should be '10 s monitoring window'.
  4. [Section IV, 'URLLC bound'] The term 'URLLC' is used for a 150 ms end-to-end bound. URLLC is typically associated with much tighter latency targets; define the service class or use a different name to avoid confusion.
  5. [Throughout] The paper would benefit from a notation table for α, β, γ, L, U, P, x, Ω, and d_t. Several symbols are introduced only in prose and used across equations without a centralized definition.

Circularity Check

2 steps flagged

The central performance claim is forced by construction: Table II's adaptive latency is the static-split latency minus improvement margins 'inherited' from cited works, so the claimed gain is the input, not a derived result; the 150 ms QoS bullet is also contradicted by the paper's own Table II.

specific steps
  1. fitted input called prediction [Section IV, 'Key Metrics' and Table II]
    "Table II summarizes expected steady-state KPIs averaged over a 10s monitoring window, where deltas inherit margins from related works [24]–[26] and scale with the analytical latency model."

    The adaptive latency column is defined as the static-split latency minus positive deltas 'inherited' from prior adaptive split-inference works [24]–[26] (with [17] for re-splitting). No measurement, simulation, or analytical model of the proposed CP/AO/SR/RB components is used to obtain these deltas. Therefore the conclusion 'Adaptive split inference thus consistently outperforms any static configuration' is guaranteed by construction: adaptive latency is static latency minus an assumed positive improvement. The predicted gain is the input, not an independently derived result.

  2. self definitional [Section IV, 'QoS guarantees' bullet; Table I; Table II]
    "QoS guarantees. The 150 ms URLLC bound is met across all loads only under the adaptive scheme, highlighting its robustness for V2X and XR services."

    The 150 ms bound is simultaneously the hand-set trigger threshold L_max in Table I and the claimed QoS guarantee. If the orchestrator reconfigures whenever EWMA latency exceeds L_max, then 'meeting 150 ms' is a design target of the control loop, not an independent prediction. Moreover, the paper's own Table II lists 200 ms adaptive latency at 20 Mb/s backhaul, which contradicts the bullet: the 150 ms bound is not met across all loads even in the paper's expected KPIs. Thus the QoS claim is neither derived from the model nor consistent with the reported numbers.

full rationale

The paper's quantitative evidence base for the central claim is circular. Section IV states that the expected KPIs are derived by combining the ETSI MEC latency model with 'empirical improvements reported by recent split inference works [15], [17]', and Table II explicitly says the deltas 'inherit margins from related works [24]–[26]'. The adaptive latency column is therefore the static latency reduced by externally supplied positive margins; the −39% to −60% latency reduction and 2.1× throughput are these margins repackaged as predictions for the proposed framework. The framework's own contributions (monitoring, adaptive orchestration, split revision, reconfiguration broadcast) are not modeled or measured. The paper does disclose that 'comprehensive cross-framework benchmarking' is future work, which mitigates the charge of hidden circularity but does not remove it: the load-bearing quantitative claim still reduces to the borrowed deltas. The 150 ms QoS bullet is additionally self-referential because L_max=150 ms is the trigger threshold, and it is contradicted by Table II's 200 ms entry at 20 Mb/s. No significant self-citation load-bearing circularity exists: the author-self citations [22], [23] are peripheral extensions, not inputs to the performance derivation. The score of 6 reflects that the central 'prediction' is constructed from its own inputs, while the architectural contribution remains independently defined.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The paper's formalizable core is a standard assignment ILP with capacity and privacy constraints; what is new is the runtime trigger/reconfiguration machinery, whose thresholds are hand-set and whose cost weights are unspecified. The main quantitative claim rests on margins imported from cited works rather than on measurement or derivation.

free parameters (8)
  • α, β, γ objective weights = not instantiated
    In the objective Φ = αL + βU + γP (Section III-B); the paper never sets or tunes them, so the optimization is not operational.
  • L_max (latency trigger) = 150 ms
    Table I 'empirical value'; hand-chosen, and reused as the QoS bound whose satisfaction is then claimed.
  • U_max (utilization trigger) = 0.85
    Table I; no justification or sensitivity analysis given.
  • B_min (bandwidth trigger) = 50 Mbps
    Table I; no justification given.
  • T_cool (reconfiguration cool-down) = 30 s
    Table I; no justification given.
  • T_proc, T_queue (latency model constants) = unspecified
    Eq. (10) requires them to produce Table II numbers; values are never stated.
  • Inherited adaptive-improvement margins from [24]–[26] = -60%/-53%/-48%/-39% latency deltas
    Section IV: 'deltas inherit margins from related works'; these external numbers convert static-split latency into adaptive-split latency — the predicted gains are these borrowed margins.
  • Expected GPU utilization (Table II) = 86–92%
    Reported with no derivation of how re-splitting yields these utilization values.
axioms (5)
  • domain assumption Latency decomposes additively as T_proc + T_queue + T_tx(bandwidth) (Eq. 10)
    Invoked as the ETSI MEC latency model without derivation or stated parameter values; it generates all absolute latency numbers in Table II.
  • domain assumption Privacy is preserved by executing sensitive layers locally
    Section III-D asserts that local execution of sensitive segments makes data reverse engineering 'significantly more challenging' (citing [6]); no threat model or quantitative analysis is given.
  • domain assumption Optimal splits are consecutive layer segments S1..Sk
    Section III-B restricts splits to consecutive segments; the optimization over Ω in Eq. (6) never explores non-consecutive or replicated placements.
  • ad hoc to paper Trigger thresholds (150 ms, 0.85, 50 Mbps, 30 s) suffice to detect when reconfiguration helps
    Table I labels these 'empirical values' with no calibration or sensitivity study; L_max doubles as the claimed QoS bound.
  • domain assumption Static splits are suboptimal under volatility
    Sections I–II motivation, standard in the cited literature; not independently established here.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Joint Partitioning and Placement of Foundation Models for Real-Time Edge AI." pith.science (2026). https://pith.science/paper/XPLA4N7D

@misc{pith2026251201039,
  author       = {Pith},
  title        = {Pith review of: Joint Partitioning and Placement of Foundation Models for Real-Time Edge AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XPLA4N7D}},
  note         = {Machine review of arXiv:2512.01039}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Inference over large-scale foundation models within heterogeneous edge environments necessitates a fundamentally reconfigurable orchestration substrate. Static partitioning of model layers presumes temporal stability across compute and network resources, which is misaligned with the volatility of real-world deployments. We introduce a framework in which both the spatial placement and internal segmentation of foundation models are elevated to runtime-resolved constructs. The orchestration problem is formalized as a constrained optimization over layer-wise assignments, subject to evolving latency, utilization, and privacy gradients. The framework implements reactive inference composition responsive to infrastructural fluctuations by integrating model-aware capacity profiling with dynamic graph re-partitioning and reallocation. We introduce architectural and algorithmic components, along with a representative use case in 6G multi-access edge computing.

Figures

Figures reproduced from arXiv: 2512.01039 by Aladin Djuhera, Alecio Binotto, Fernando Koch.

Figure 1
Figure 1. Figure 1: DSI with LFMs requires adaptive orchestration between edge and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Reference architecture of the proposed adaptive split inference orchestration. Sub-split models (S1, S2, S3) are deployed across edge/cloud nodes, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Expected latency vs. bandwidth availability in our 5G MEC urban [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 12 linked inside Pith

  1. [1]

    Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities,

    H. Zhou, C. Hu, Y . Yuan, Y . Cui, Y . Jin, C. Chen, H. Wu, D. Yuan, L. Jiang, D. Wu, X. Liu, C. Zhang, X. Wang, and J. Liu, “Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities,”IEEE Com- munications Surveys & Tutorials, pp. 1–1, 2024

    2024

  2. [2]

    LLM Inference Serv- ing: Survey of Recent Advances and Opportunities,

    B. Li, Y . Jiang, V . Gadepally, and D. Tiwari, “LLM Inference Serv- ing: Survey of Recent Advances and Opportunities,”arXiv preprint arXiv:2407.12391, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Split computing: DNN Inference Partition with Load Balancing in IoT-Edge Platform for Beyond 5G,

    J. Karjee, P. Naik S, K. Anand, and V . N. Bhargav, “Split computing: DNN Inference Partition with Load Balancing in IoT-Edge Platform for Beyond 5G,”Measurement: Sensors, vol. 23, p. 100409, 2022

    2022

  4. [4]

    Split Computing: Dynamic Partitioning and Reliable Communications in IoT-Edge for 6G Vision,

    J. Karjee, K. Anand, V . N. Bhargav, P. S. Naik, R. B. V . Dabbiru, and N. Srinidhi, “Split Computing: Dynamic Partitioning and Reliable Communications in IoT-Edge for 6G Vision,” in2021 8th International Conference on Future Internet of Things and Cloud (FiCloud), 2021

    2021

  5. [5]

    Kubernetes and Docker Load Balancing: State-of-the-Art Techniques and Challenges,

    I. Vasireddy, G. Ramya, and P. Kandi, “Kubernetes and Docker Load Balancing: State-of-the-Art Techniques and Challenges,”International Journal of Innovative Research in Engineering and Management, vol. 10, no. 6, pp. 49–54, 2023

    2023

  6. [6]

    Privacy-Preserving Machine Learning: Methods, Challenges and Directions,

    R. Xu, N. Baracaldo, and J. Joshi, “Privacy-Preserving Machine Learning: Methods, Challenges and Directions,”arXiv preprint arXiv:2108.04417, 2021

    Pith/arXiv arXiv 2021

  7. [7]

    Edge Artificial Intelligence for 6G: Vision, Enabling Technologies, and Applications,

    K. B. Letaief, Y . Shi, J. Lu, and J. Lu, “Edge Artificial Intelligence for 6G: Vision, Enabling Technologies, and Applications,”IEEE journal on selected areas in communications, vol. 40, no. 1, pp. 5–36, 2021

    2021

  8. [8]

    Kubernetes Scheduling: Taxonomy, Ongoing Issues and Challenges,

    C. Carri ´on, “Kubernetes Scheduling: Taxonomy, Ongoing Issues and Challenges,”ACM Computing Surveys, vol. 55, no. 7, pp. 1–37, 2022

    2022

  9. [9]

    Efficient Training of Large Language Models on Distributed Infrastructures: A Survey,

    J. Duan, S. Zhang, Z. Wanget al., “Efficient Training of Large Language Models on Distributed Infrastructures: A Survey,” 2024. [Online]. Available: https://arxiv.org/abs/2407.20018

    Pith/arXiv arXiv 2024

  10. [10]

    A Comprehensive Study on Quantiza- tion Techniques for Large Language Models,

    J. Lang, Z. Guo, and S. Huang, “A Comprehensive Study on Quantiza- tion Techniques for Large Language Models,” inIEEE ICAIRC, 2024

    2024

  11. [11]

    A Survey on Efficient Inference for Large Language Models,

    Z. Zhou, X. Ning, K. Hong, T. Fu, J. Xu, S. Li, Y . Lou, L. Wang, Z. Yuan, X. Liet al., “A Survey on Efficient Inference for Large Language Models,”arXiv preprint arXiv:2404.14294, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Artificial General Intelligence (AGI)- Native Wireless Systems: A Journey Beyond 6G,

    W. Saad, O. Hashash, C. K. Thomas, C. Chaccour, M. Debbah, N. Mandayam, and Z. Han, “Artificial General Intelligence (AGI)- Native Wireless Systems: A Journey Beyond 6G,” 2024. [Online]. Available: https://arxiv.org/abs/2405.02336

    Pith/arXiv arXiv 2024

  13. [13]

    Dis- tributed Inference Acceleration with Adaptive DNN Partitioning and Offloading,

    T. Mohammed, C. Joe-Wong, R. Babbar, and M. D. Francesco, “Dis- tributed Inference Acceleration with Adaptive DNN Partitioning and Offloading,” inIEEE INFOCOM 2020 - IEEE Conference on Computer Communications, 2020, pp. 854–863

    2020

  14. [14]

    Pushing Large Language Models to the 6G Edge: Vision, Challenges, and Opportunities,

    Z. Lin, G. Qu, Q. Chen, X. Chen, Z. Chen, and K. Huang, “Pushing Large Language Models to the 6G Edge: Vision, Challenges, and Opportunities,”arXiv preprint arXiv:2309.16739, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    EdgeShard: Efficient LLM Inference via Collaborative Edge Computing,

    M. Zhang, X. Shen, J. Cao, Z. Cui, and S. Jiang, “EdgeShard: Efficient LLM Inference via Collaborative Edge Computing,”IEEE Internet of Things Journal, pp. 1–1, 2024

    2024

  16. [16]

    QoS-Aware Edge AI Placement and Scheduling with Multiple Implementations in FaaS-Based Edge Computing,

    N. Hudson, H. Khamfroush, M. Baughman, D. E. Lucani, K. Chard, and I. Foster, “QoS-Aware Edge AI Placement and Scheduling with Multiple Implementations in FaaS-Based Edge Computing,”Future Generation Computer Systems, vol. 157, pp. 250–263, 2024

    2024

  17. [17]

    Adaptive Layer Splitting for Wireless LLM Inference in Edge Computing: A Model- Based Reinforcement Learning Approach,

    Y . Chen, R. Li, X. Yu, Z. Zhao, and H. Zhang, “Adaptive Layer Splitting for Wireless LLM Inference in Edge Computing: A Model- Based Reinforcement Learning Approach,” 2024. [Online]. Available: https://arxiv.org/abs/2406.02616

    Pith/arXiv arXiv 2024

  18. [18]

    Distributing Deep Neural Networks with Containerized Partitions at the Edge,

    L. Zhou, H. Wen, R. Teodorescu, and D. H. Du, “Distributing Deep Neural Networks with Containerized Partitions at the Edge,” in2nd USENIX Workshop on Hot Topics in Edge Computing, 2019

    2019

  19. [19]

    Learning- Aided Computation Offloading for Trusted Collaborative Mobile Edge Computing,

    Y . Li, X. Wang, X. Gan, H. Jin, L. Fu, and X. Wang, “Learning- Aided Computation Offloading for Trusted Collaborative Mobile Edge Computing,”IEEE Transactions on Mobile Computing, vol. 19, no. 12, pp. 2833–2849, 2019

    2019

  20. [20]

    Optimal AI Model Splitting and Resource Allocation for Device-Edge Co-Inference in Multi-User Wireless Sensing Systems,

    X. Li and S. Bi, “Optimal AI Model Splitting and Resource Allocation for Device-Edge Co-Inference in Multi-User Wireless Sensing Systems,” IEEE Transactions on Wireless Communications, vol. 23, no. 9, pp. 11 094–11 108, 2024

    2024

  21. [21]

    Optimum Splitting Comput- ing for DNN Training Through Next Generation Smart Networks: A Multi-Tier Deep Reinforcement Learning Approach,

    S.-Y . Lien, C.-H. Yeh, and D.-J. Deng, “Optimum Splitting Comput- ing for DNN Training Through Next Generation Smart Networks: A Multi-Tier Deep Reinforcement Learning Approach,”Wireless Networks, vol. 30, no. 3, pp. 1737–1751, 2024

    2024

  22. [22]

    R-SFLLM: Jamming Resilient Framework for Split Federated Learning with Large Language Models,

    A. Djuhera, V . C. Andrei, X. Li, U. J. M ¨onich, H. Boche, and W. Saad, “R-SFLLM: Jamming Resilient Framework for Split Federated Learning with Large Language Models,” 2024. [Online]. Available: https://arxiv.org/abs/2407.11654

    arXiv 2024

  23. [23]

    SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging,

    A. Djuhera, S. R. Kadhe, F. Ahmed, S. Zawad, and H. Boche, “SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging,” 2025. [Online]. Available: https://arxiv.org/abs/2503.17239

    Pith/arXiv arXiv 2025

  24. [24]

    AMP4EC: Adaptive Model Partitioning Framework for Efficient Deep Learning Inference in Edge Computing Environments,

    G. Zhang, W. Guo, Z. Tan, and H. Jiang, “AMP4EC: Adaptive Model Partitioning Framework for Efficient Deep Learning Inference in Edge Computing Environments,” 2025. [Online]. Available: https://arxiv.org/abs/2504.00407

    Pith/arXiv arXiv 2025

  25. [25]

    SplitPlace: AI Augmented Splitting and Placement of Large-Scale Neural Networks in Mobile Edge Environments,

    S. Tuli, G. Casale, and N. R. Jennings, “SplitPlace: AI Augmented Splitting and Placement of Large-Scale Neural Networks in Mobile Edge Environments,” 2022. [Online]. Available: https://arxiv.org/abs/2205.10635

    Pith/arXiv arXiv 2022

  26. [26]

    Adaptive Compression-Aware Split Learning and Inference for Enhanced Network Efficiency,

    A. Mudvari, A. Vainio, I. Ofeidis, S. Tarkoma, and L. Tassiulas, “Adaptive Compression-Aware Split Learning and Inference for Enhanced Network Efficiency,” 2024. [Online]. Available: https://arxiv.org/abs/2311.05739

    Pith/arXiv arXiv 2024

  27. [27]

    The Llama 3 Herd of Models,

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The Llama 3 Herd of Models,”arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.