Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Pruning-Aware Multi-Cluster Co-Inference for Large AI Models in AI-RANs

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that jointly optimizing per-device pruning ratios, GPU task scheduling, bandwidth, and transmit power minimizes inference distortion in multi-cluster edge co-inference, guided by rate-distortion and…

desk verdict A plausible multi-cluster co-inference formulation with a real system-level contribution, but the load-bearing distortion surrogate is mis-derived and never validated against measured accuracy, so the central outperformance claim is conditional. read the letter →

arxiv 2608.03026 v1 pith:RSQO6OPD submitted 2026-08-04 cs.DC

Xiaowen Cao , Zhonghao Lyu , Shicheng Chu , Zezhong Zhang , Dingzhu Wen , Guangxu Zhu , Kaibin Huang , Shuguang Cui
show 1 more author
Jie Xu
This is my paper · ORCID
classification cs.DC
keywords largeAImodelsco-inferencemodelpruningrate-distortiontheorypartialinformationdecompositionShapleyvalueGPUschedulingresourceallocation
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 proposes a multi-cluster edge co-inference design in which devices extract local features with pruned models, upload them, and a GPU-equipped server fuses the features and completes inference. It tries to establish that jointly choosing each device's pruning ratio, GPU queue assignment, bandwidth, and transmit power—guided by how much each device's view contributes to the task—minimizes inference distortion under latency and energy budgets. The authors derive an importance-weighted rate-distortion distortion estimate and use partial information decomposition, in the form of Shapley values, to rank device contribution. Simulations on image classification and video-to-text captioning show the joint design beating fixed-power, fixed-bandwidth, round-robin, random-scheduling, and equal-importance baselines. If the argument is right, it gives a principled way to deploy large AI models in spectrum- and GPU-constrained edge networks.

What carries the argument

The load-bearing object is the importance-weighted distortion-rate function $\hat{D}_{m,k}(\rho_{m,k})$ in Eq. (23). It turns the paper's qualitative trade-off—pruning saves computation and communication but hurts inference—into a per-device numerical cost: a device with larger Shapley-based importance $\alpha_{m,k}$ pays a larger weighted distortion for the same pruning ratio, so the optimizer prunes it less. That surrogate is what couples pruning decisions to bandwidth, power, and GPU queueing in the joint problem (P1), and it is also what makes the claim about uniform pruning being suboptimal testable.

What would settle it

Concretely, on a held-out multi-view task, compute per-device measured distortion (e.g., drop in task accuracy or CIDEr) for a grid of pruning-ratio vectors and compare the ordering predicted by Eq. (23). Any inversion—a device ranked low-importance that can be pruned hard without hurting measured performance, or a high-importance device that degrades despite light pruning—would refute the surrogate and collapse the joint optimization's claimed gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that model compression and communication/computation resource allocation in collaborative edge inference should be treated as one coupled problem whose objective is an importance-weighted pruning distortion. It derives a rate-distortion lower bound for pruned LAIM parameters (Eq. (22)) and converts it into a per-device distortion estimate $\hat{D}_{m,k}(\rho_{m,k})$ (Eq. (23)) scaled by an importance coefficient $\alpha_{m,k}$ obtained from Shapley-value approximation of partial information decomposition. Jointly minimizing the sum of these distortions over pruning ratios, GPU scheduling, bandwidth, and transmit power, subject to delay and energy constraints, makes high-contribution devices keep more parameters while low-contribution devices absorb the compression. The numerical evaluation on CIFAR-10 and MSR-VTT reports higher classification accuracy and captioning scores (BLEU-4, CIDEr) at the same resource budgets, and latency-decomposition experiments show the scheduling decisions keeping heterogeneous server workloads inside the deadline.

Load-bearing premise

The optimization's accuracy gains all flow from the assumption that the weighted-distortion formula $\hat{D}_{m,k}(\rho_{m,k})$ correctly ranks which devices can be pruned hardest in a real model, and that ranking is never validated against measured task error.

Editorial extensions

If this is right

  • A device with larger Shapley/PID importance should be pruned less, and uniform pruning across devices is generally suboptimal.
  • When no pruning is needed, importance no longer steers bandwidth; the pruning point gates whether semantic importance affects resource decisions.
  • Joint GPU scheduling plus communication optimization keeps heterogeneous server workloads within the latency threshold, while round-robin and random scheduling miss it.
  • The same design improves image classification accuracy and video-to-text captioning quality (BLEU-4, CIDEr) at the same energy and latency budgets, showing the framework is not tied to one task metric.

Reading between the lines

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

  • Editorial inference: the validity of Eq. (23) as a ranking device could be checked against measured output distortion on larger, non-Laplacian models; that check would separate the rate-distortion theory from the surrogate's assumptions.
  • Editorial inference: Shapley coefficients estimated from labeled evaluation data may drift when sensing conditions change; periodic re-estimation or an online update rule would be a natural extension.
  • Editorial inference: the gate mechanism, where saturation of pruning blocks importance from influencing bandwidth, likely generalizes to quantization and knowledge distillation choices as well.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a multi-cluster LAIM co-inference framework in which an edge server with multiple GPUs coordinates several user clusters, each performing multi-view feature extraction with pruned on-device models and uploading the features for server-side aggregation and GPU-scheduled inference. The authors derive a rate-distortion lower bound for single-device pruning, extend it to a multi-device importance-weighted bound using partial information decomposition and Shapley-value-based importance coefficients, and formulate a joint optimization problem (P1) over pruning ratios, task scheduling, bandwidth allocation, and transmit power under latency and energy constraints. The problem is solved by alternating between an MILP for GPU scheduling and an SCA-based convex approximation for resource allocation. The paper evaluates the approach on CIFAR-10 image classification and MSR-VTT video-to-text captioning, reporting accuracy, BLEU-4, CIDEr, and latency-decomposition results against five benchmark schemes. The central claim is that the proposed importance-aware joint design significantly outperforms the benchmarks and that the rate-distortion/PID analysis characterizes the fundamental pruning-inference trade-off.

Significance. If the presented surrogate distortion correctly ranks pruning choices and the optimization is correctly derived, the framework would be a useful contribution to edge AI resource allocation, extending single-user pruning-aware co-inference to a multi-cluster setting with GPU queueing, bandwidth, and power control. The problem formulation is ambitious and the solution decomposition into MILP and SCA is coherent and appropriate. The experimental scope is broader than many papers in this area, covering two task families and reporting both threshold sweeps and task-level latency decompositions. However, the paper's central quantitative claim rests on the importance-weighted distortion surrogate in Eq. (23), and that surrogate currently has algebraic inconsistencies with the stated rate-distortion bound and is not empirically calibrated against measured task distortion. The Shapley-value identification in Eq. (28) is asserted rather than proven, and the numerical curves lack error bars or seed information.

major comments (3)
  1. [§III, Proposition 1 and Eqs. (22)–(23)] The derivation of the surrogate distortion is algebraically inconsistent with Lemma 2. Substituting D*_{m,k} = q_{m,k} D / (Qtot_m α_{m,k}) into the single-device bound (21) gives Rdot_{m,k}(D_{m,k}) ≥ h(W^{(m,k)}) + q_{m,k} log(Qtot_m α_{m,k} / (√(πe) D)) + log(Γ(q_{m,k}/2)/(2Γ(q_{m,k}))). Eq. (22) instead prints q_{m,k} log(Qtot_m α_{m,k} √(πe) D), which increases with D and directly contradicts the trade-off stated in Remark 2. Eq. (23) then prints α_{m,k} q_{m,k} √(πe) 2^{-(...)}, dropping the Qtot_m factor, changing the Gamma normalization, and using q_{m,k} where Qtot_m appears in the substitution. Since Dhat_{m,k}(ρ_{m,k}) is the objective of problem (P1), the central optimization is currently not derived from the stated rate-distortion bound. The authors should correct Eqs. (22)–(23) with the full substitution, verify the monotonicity of the corrected expression, and re-run the optimization to check whether the qualitative conclusions survive.
  2. [§III.3, Eq. (28)] The identification of the importance coefficient α_{m,k} with a Shapley value is asserted rather than proven. In Proposition 1, α_{m,k} is a Lagrange multiplier weight in a synthetic weighted-distortion constraint, whereas Eq. (28) computes a Shapley value of predictive log-loss coalition values on labeled evaluation data. These are not the same object, and no argument is given that the Shapley attribution recovers the α that makes the rate-distortion bound tight or that correctly ranks pruning choices. This matters because the entire importance-aware pruning and bandwidth-allocation behavior in Figs. 3–5 and Fig. 8 depends on this identification. The authors should either prove an equivalence under stated assumptions or treat the Shapley-based α as an empirical proxy and validate it directly by comparing surrogate-predicted pruning rankings with measured accuracy, BLEU-4, and CIDEr on held-out data.
  3. [§V, Figs. 3–5] The numerical results do not support the word 'significantly' in the abstract and conclusions. None of the accuracy, BLEU-4, or CIDEr curves report error bars, confidence intervals, or the number of random seeds/trials, so the reader cannot assess whether the proposed design's margins over the benchmarks are statistically meaningful. Moreover, the paper never plots the surrogate distortion Dhat against the measured task metrics; without such a calibration, the experiments cannot distinguish a correct surrogate from a surrogate whose ranking errors are masked by the resource-allocation gains. I would ask for repeated-seed results with error bars and a direct surrogate-vs-measured-distortion calibration plot, at least for the CIFAR-10 and video-captioning operating points used in the threshold sweeps.
minor comments (5)
  1. [§III, Proposition 1 proof] The constant term in the Lagrangian (25) is printed as q_{m,k} log(q_{m,k} √(πe)), but the preceding expression in (21) implies q_{m,k} log(q_{m,k} / √(πe)); the sign of the √(πe) term should be corrected for consistency.
  2. [§II.C, Eq. (8)] The text has a formatting artifact 'the delay for on-device inference is 1 tI_{m,k}(ρ_{m,k}) = ...'; the stray '1' should be removed.
  3. [§III, Eq. (23)] The relation between the rate R_{m,k} and the pruning ratio ρ_{m,k}, namely R_{m,k} = ρ_{m,k} q_{m,k} γ, is introduced only implicitly in Eq. (23) after γ is defined as quantization bits per parameter; the authors should state this relation explicitly before using it in the distortion-rate expression.
  4. [§V, Footnote 3] The footnote acknowledges that the experimental models are relatively lightweight and that memory capacity, KV-cache management, and data retrieval are not modeled; this is an important scope limitation for the 'large AI model' claim, and the paper should discuss explicitly how the proposed framework transfers to billion-parameter models with memory-bound inference.
  5. [§III.3, Footnote 2] The footnote states that importance coefficients are estimated offline or periodically from historical data and are not part of the co-inference latency; this is a sensible choice, but the paper should also state how sensitive the reported gains are to the freshness or size of the evaluation dataset used for Shapley estimation.

Circularity Check

2 steps flagged · score 5.0 of 10

The analytical trade-off claim rests on a load-bearing self-citation and an importance coefficient that is constructed to produce its own conclusion, though the empirical benchmark comparisons retain independent content.

  1. self citation load bearing [Section III, Lemma 2, Eq. (21); reused in Proposition 1, Eqs. (22)-(23), and in the objective of (P1)]
    "Lemma 2. Assume that the LAIM parameters W ∈ Rq follow a multivariate Laplacian distribution. Under the distortion function ‖W − Ŵ‖F ≤ D, from [12], a lower bound of the rate-distortion function is ˙R(D) ≥ h(W ) + q log(q√πeD) + log(Γ(q/2)/2Γ(q)), where h(W ) denotes the differential entropy of W."

    Equation (21) is the sole basis for the multi-device bound in Proposition 1 and for the surrogate objective Dhat in Eq. (23) that is minimized in (P1). It is imported verbatim from [12], whose author list overlaps with the present paper (Z. Lyu and J. Xu), and no derivation, proof sketch, or independent verification is provided in this manuscript. The paper's claimed 'rate-distortion-based analytical framework' thus inherits its central mathematical content from the authors' own prior work, making the analytical contribution load-bearing on a self-citation rather than on an independently established or re-derived result.

  2. self definitional [Section III-2 and III-3, Proposition 1 proof, Remark 3, and Eq. (28)]
    "By setting the derivative of (25) w.r.t. Dm,k as zero, i.e., ∂L/∂Dm,k = − qm,k/Dm,k + λαm,k = 0, we have D⋆m,k = qm,k/λαm,k. ... a device with a larger importance coefficient incurs a larger weighted distortion penalty. Therefore, important devices require more preserved model parameters."

    The coefficient α is introduced as the weight in the artificial importance-weighted constraint ∑αm,kDm,k ≤ D, and the KKT condition immediately makes D⋆m,k inversely proportional to αm,k. The Remark's conclusion that high-α devices need lower pruning is therefore an algebraic restatement of how α was defined, not an empirical or PID-derived prediction. The later identification of α with the Shapley value of predictive log-loss in Eq. (28) is asserted rather than derived from the rate-distortion model, so the 'importance-aware' trade-off is partly constructed by definition rather than characterized independently.

full rationale

The paper's empirical contribution—joint optimization of pruning ratios, task scheduling, bandwidth, and power evaluated on CIFAR-10 and MSR-VTT against five baselines—is not itself circular: the reported accuracy, BLEU-4, and CIDEr numbers are external to the surrogate objective Dhat, and the Shapley coefficients are described as estimated offline from representative training or historical data rather than fitted to the test set. However, the paper's theoretical wrapper has two load-bearing circularity concerns. First, Lemma 2, the single-device rate-distortion lower bound that feeds Proposition 1 and the entire Dhat objective in (P1), is taken without derivation from [12], a prior paper by the present authors; no independent proof or machine-checked support is supplied. Second, the claim that important devices should be pruned less follows directly from defining α as the weight in the weighted-distortion constraint, so the 'insight' is a property of the construction; the leap from that Lagrangian weight to a Shapley/PID contribution in Eq. (28) is asserted, not proven. There is also an internal algebraic inconsistency in the printed Eq. (21)/(22), where the rate bound increases with D, contrary to both the proof's Lagrangian and Remark 2; this is a correctness defect that further weakens the analytical derivation. These issues do not force the empirical benchmark gains, but they mean the paper's central 'fundamental trade-off' claim is substantially built on a self-citation plus a definitional coefficient, warranting a moderate circularity score rather than a clean bill.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The main load-bearing elements are imported from the same author group's prior rate-distortion work, plus the empirical Shapley importance coefficients and the linear pruning workload model. No new physical entities are introduced. The key unvalidated step is treating the importance-weighted parameter distortion as a faithful surrogate for task-level inference distortion.

free parameters (1)
  • Importance coefficients alpha_{m,k} = Estimated from labeled evaluation data via Monte Carlo Shapley (Eq. 28)
    Central to the importance-weighted distortion objective; measured offline from a dataset with ground-truth labels, not derived from first principles.
assumptions (6)
  • domain assumption Output distortion is upper bounded by parameter distortion with a Lipschitz constant (Lemma 1)
    Imported from [12]; requires smoothness of activations, no validation for ViT-Tiny or TinyViT/Qwen3 in this paper.
  • domain assumption LAIM parameters follow a multivariate Laplacian distribution, yielding the rate-distortion lower bound in Lemma 2
    Imported from [12]; used to derive Eq. (23). The only support here is Fig. 2, which claims Laplacian-looking weight histograms.
  • ad hoc to paper On-device computation delay and energy scale linearly with pruning ratio rho (Eq. 8 and Eq. 16)
    The author footnote explicitly calls this an analytical approximation rather than a universal hardware law.
  • ad hoc to paper Shapley value computed on labeled historical data is a valid proxy for PID device contribution at inference time
    No proof or reference that Shapley recovers PID structure beyond two devices; footnote 2 says coefficients are reused across tasks.
  • ad hoc to paper The global importance-weighted distortion constraint sum alpha D <= D captures task-level inference distortion
    This is the key modeling step in Proposition 1, with no empirical validation against actual task distortion.
  • domain assumption Clusters are task-isolated and static; wireless channels, device clusters, and GPU queues are known and stationary
    System model in Sec. II; the conclusion lists dynamic environments as future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pruning-Aware Multi-Cluster Co-Inference for Large AI Models in AI-RANs." pith.science (2026). https://pith.science/paper/RSQO6OPD

@misc{pith2026260803026,
  author       = {Pith},
  title        = {Pith review of: Pruning-Aware Multi-Cluster Co-Inference for Large AI Models in AI-RANs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RSQO6OPD}},
  note         = {Machine review of arXiv:2608.03026}
}
read the original abstract

The increasing scale and computational demands of large artificial intelligence models (LAIMs) present significant challenges for efficient inference in resource-constrained distributed environments. In this paper, we propose a multi-cluster LAIM co-inference framework, where an edge server equipped with multiple graphics processing units (GPUs) coordinates multiple user clusters to execute inference tasks collaboratively. Within each cluster, devices capture data from diverse perspectives and employ lightweight on-device LAIMs to extract local features. These features are then transmitted to the edge server, where they are aggregated and fused to generate a more accurate inference outcome. To reveal the fundamental trade-off between model pruning and collaborative inference performance, we develop a theoretical framework that characterizes the impact of pruning ratios and device contributions using rate-distortion theory and partial information decomposition. Based on this analysis, we formulate a joint optimization problem that determines the model pruning ratio, the task scheduling strategy, the bandwidth allocation, and the transmission power, with the goal of minimizing the inference distortion while satisfying the constraints of latency, energy consumption, and server capacity. Extensive simulation results demonstrate that the proposed framework significantly outperforms existing benchmark schemes, achieving superior inference accuracy and resource efficiency in multi-cluster edge intelligence networks.

Figures

Figures reproduced from arXiv: 2608.03026 by the authors.

Figure 1
Figure 1. Illustration of multi-cluster LAIM co-inference [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Laplacian distributed weights of ResNet-152 and G [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Classification accuracy on CIFAR-10 under differe [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: BLEU-4 scores for the video-to-text captioning ta [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: CIDEr scores for the video-to-text captioning tas [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Task-level inference latency decomposition for t [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Task-level inference latency decomposition for t [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Device importance coefficients, retained-model r [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages

  1. [1]

    AI-native O-RA N archi- tectures for 6G: Toward real-time adaptation, conflict reso lution, and efficient resource management,

    S. Salmi, M. A. Ouameur, M. Bagaa, G. C. Alexandropoulos, A. Tahenni, D. Massicotte, and A. Ksentini, “AI-native O-RA N archi- tectures for 6G: Toward real-time adaptation, conflict reso lution, and efficient resource management,” IEEE Transactions on Network and Service Management, vol. 23, pp. 4110–4121, 2026

    work page 2026

  2. [2]

    A Systematic Perspective on Co mmuni- cation Innovations Toward 6G,

    N. Ge, C. Jiang, and J. Lu, “A Systematic Perspective on Co mmuni- cation Innovations Toward 6G,” Engineering, vol. 31, pp. 19–24, Dec. 2023

    work page 2023

  3. [3]

    Joint sensing, communication, and computation for vertical fede rated edge learning in edge perception networks,

    X. Cao, D. Wen, S. Bi, Y . Cui, G. Zhu, H. Hu, and Y . C. Eldar, “ Joint sensing, communication, and computation for vertical fede rated edge learning in edge perception networks,” IEEE Trans. Mobile Comput. , pp. 1–14, 2026

    work page 2026

  4. [4]

    Vision-language mo dels for vision tasks: A survey,

    J. Zhang, J. Huang, S. Jin, and S. Lu, “Vision-language mo dels for vision tasks: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5625–5644, 2024

    work page 2024

  5. [5]

    Lambo: Large AI model empowered edge intelligence,

    L. Dong, F. Jiang, Y . Peng, K. Wang, K. Y ang, C. Pan, and R. S chober, “Lambo: Large AI model empowered edge intelligence,” IEEE Commun. Mag., vol. 63, no. 4, pp. 88–94, 2025

    work page 2025

  6. [6]

    Sense Smarter, Think Better: A Survey on Edge Perception for Next-Generation Networks

    Z. Lyu, X. Cao, X. Song, Y . Li, J. Wang, Y . Cui, W. Y uan, X. Y u , G. Zhu, J. Xu, D. W. K. Ng, D. Niyato, and S. Cui, “Sense smarter , think better: Edge perception for next-generation network s,” 2026. [Online]. Available: https://arxiv.org/abs/2605.18457

    work page Pith review arXiv 2026

  7. [7]

    Key Challenges and Research Directions for Space-Air-Ground Integrated Emergency Com munication Networks,

    B. Xu, H. Zhao, J. Kang, and D. Niyato, “Key Challenges and Research Directions for Space-Air-Ground Integrated Emergency Com munication Networks,” Engineering, vol. 54, pp. 52–55, Nov. 2025

    work page 2025

  8. [8]

    Mobil e edge intelligence for large language models: A contemporary sur vey,

    G. Qu, Q. Chen, W. Wei, Z. Lin, X. Chen, and K. Huang, “Mobil e edge intelligence for large language models: A contemporary sur vey,” IEEE Commun. Surveys Tuts. , vol. 27, no. 6, pp. 3820–3860, 2025

    work page 2025

Show all 40 references
  1. [9]

    Closing the generalization gap in parameter-efficient federated ed ge learning,

    X. Du, Z. Lyu, X. Cao, C. Wen, S. Cui, and J. Xu, “Closing the generalization gap in parameter-efficient federated ed ge learning,”

  2. [10]

    Gemini: a family of highly capable multimodal models,

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Y u, R. So ricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican et al., “Gemini: a family of highly capable multimodal models,” arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  3. [11]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P . Albert, A. Almahairi , Y . Babaei, N. Bashlykov, S. Batra, P . Bhargava, S. Bhosale et al. , “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  4. [12]

    The l arger the merrier? Efficient Large AI model inference in wireless edge networks,

    Z. Lyu, M. Xiao, J. Xu, M. Skoglund, and M. D. Renzo, “The l arger the merrier? Efficient Large AI model inference in wireless edge networks,” IEEE J. Sel. Areas Commun. , pp. 1–1, 2025

    2025

  5. [13]

    Optimal AI model splitting and resource allocation for device-edge co-inference in multi-user wireless sensing s ystems,

    X. Li and S. Bi, “Optimal AI model splitting and resource allocation for device-edge co-inference in multi-user wireless sensing s ystems,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 11 094–11 108, Sep. 2024

    2024

  6. [14]

    Large language m odels (llms) inference offloading and resource allocation in clou d-edge com- puting: An active inference approach,

    Y . He, J. Fang, F. R. Y u, and V . C. Leung, “Large language m odels (llms) inference offloading and resource allocation in clou d-edge com- puting: An active inference approach,” IEEE Transactions on Mobile Computing, vol. 23, no. 12, pp. 11 253–11 264, 2024

    2024

  7. [15]

    Mcg-sched: Mu lti- cluster gpu scheduling for resource fragmentation reducti on and load balancing,

    H. Wu, X. Wang, X. Luo, W. Shen, and W. Lin, “Mcg-sched: Mu lti- cluster gpu scheduling for resource fragmentation reducti on and load balancing,” IEEE Transactions on Parallel and Distributed Systems , vol. 36, no. 12, pp. 2789–2800, 2025

    2025

  8. [16]

    Semantic communi cations for image recovery and classification via deep joint source a nd channel coding,

    Z. Lyu, G. Zhu, J. Xu, B. Ai, and S. Cui, “Semantic communi cations for image recovery and classification via deep joint source a nd channel coding,” IEEE Trans. Wireless Commun. , vol. 23, no. 8, pp. 8388–8404, 2024

    2024

  9. [17]

    Semantic communi cation: A survey on research landscape, challenges, and future direc tions,

    T. M. Getu, G. Kaddoum, and M. Bennis, “Semantic communi cation: A survey on research landscape, challenges, and future direc tions,” Proc. IEEE, vol. 112, no. 11, pp. 1649–1685, 2024

    2024

  10. [18]

    Adaptable semanti c compression and resource allocation for task-oriented communications ,

    C. Liu, C. Guo, Y . Y ang, and N. Jiang, “Adaptable semanti c compression and resource allocation for task-oriented communications ,” IEEE Trans. Cog. Commun. Netw. , vol. 10, no. 3, pp. 769–782, June 2024

    2024

  11. [19]

    Semantic communication meets edge intelligence : Semantic- relay-aided text transmissions,

    Z. Hu, C. Y ou, T. Liu, D. Wen, Y . Hu, Y . Cui, Y . Gong, and K. Huang, “Semantic communication meets edge intelligence : Semantic- relay-aided text transmissions,” IEEE Internet Things J. , vol. 11, no. 24, pp. 39 409–39 423, Dec 2024

    2024

  12. [20]

    Optimal model placeme nt and online model splitting for device-edge co-inference,

    J. Y an, S. Bi, and Y .-J. A. Zhang, “Optimal model placeme nt and online model splitting for device-edge co-inference,” IEEE Trans. Wireless Commun., vol. 21, no. 10, pp. 8354–8367, 2022

    2022

  13. [21]

    Task- oriented sensing, computation, and communication integra tion for multi- device edge AI,

    D. Wen, P . Liu, G. Zhu, Y . Shi, J. Xu, Y . C. Eldar, and S. Cui , “Task- oriented sensing, computation, and communication integra tion for multi- device edge AI,” IEEE Transac. Wireless Commun. , vol. 23, no. 3, pp. 2486–2502, 2024

    2024

  14. [22]

    Beyond the cloud: Edge inference for generativ e large language models in wireless networks,

    X. Zhang, J. Nie, Y . Huang, G. Xie, Z. Xiong, J. Liu, D. Niy ato, and X. Shen, “Beyond the cloud: Edge inference for generativ e large language models in wireless networks,” IEEE Trans. Wireless Commun., vol. 24, no. 1, pp. 643–658, 2025

    2025

  15. [23]

    Adaptive lay er splitting for wireless LLM inference in edge computing: A mo del- based reinforcement learning approach,

    Y . Chen, R. Li, X. Y u, Z. Zhao, and H. Zhang, “Adaptive lay er splitting for wireless LLM inference in edge computing: A mo del- based reinforcement learning approach,” 2024. [Online]. A vailable: https://arxiv.org/abs/2406.02616

    2024 arXiv

  16. [24]

    EdgeShar d: Efficient LLM inference via collaborative edge computing,

    M. Zhang, X. Shen, J. Cao, Z. Cui, and S. Jiang, “EdgeShar d: Efficient LLM inference via collaborative edge computing,” IEEE Internet Things J., vol. 12, no. 10, pp. 13 119–13 131, May 2025

    2025

  17. [25]

    Joint inference offloading and model caching for small and large language mod el collaboration,

    X. Xu, G. Feng, Y . Liu, S. Qin, J. Wang, and Y . Wang, “Joint inference offloading and model caching for small and large language mod el collaboration,” IEEE Transactions on Mobile Computing , vol. 25, no. 2, pp. 2691–2706, Feb 2026

    2026

  18. [26]

    Quantization-aware collaborative inference for large em bodied ai mod- els,

    Z. Lyu, M. Xiao, M. Skoglund, M. Debbah, and H. V . Poor, “Quantization-aware collaborative inference for large em bodied ai mod- els,” arXiv preprint arXiv:2602.13052 , 2026

    2026

  19. [27]

    Jppo++: Join t power and denoising-inspired prompt optimization for mobile llm ser vices,

    F. Y ou, H. Du, K. Huang, and A. Jamalipour, “Jppo++: Join t power and denoising-inspired prompt optimization for mobile llm ser vices,” IEEE Trans. Mobile Comput. , pp. 1–16, 2026

    2026

  20. [28]

    Berger, Rate-distortion theory

    T. Berger, Rate-distortion theory . Wiley Encyclopedia of Telecommu- nications, 2003

    2003

  21. [29]

    L. S. Shapley, A V alue for N-Person Games. Santa Monica, CA: RAND Corporation, 1952

    1952

  22. [30]

    Data shapley: Equitable valuat ion of data for machine learning,

    A. Ghorbani and J. Zou, “Data shapley: Equitable valuat ion of data for machine learning,” in Proc. Int. Conf. Mach. Learn. (ICML) , ser. Proc. Mach. Learn. Res., vol. 97. PMLR, 2019, pp. 2242–2251

    2019

  23. [31]

    Optimization

    G. Optimization. LLC, 2021. [Online]. Available: http s://www.gurobi. com

    2021

  24. [32]

    Boyd and L

    S. Boyd and L. V andenberghe, Convex Optimization,. Cambridge Univ. Press, 2004

    2004

  25. [33]

    CVX: Matlab software for discipli ned convex programming,

    M. Grant and S. Boyd, “CVX: Matlab software for discipli ned convex programming,” 2016. [Online]. Available: http://cvxr.co m/cvx

    2016

  26. [34]

    Task-oriented communica tion for mul- tidevice cooperative edge inference,

    J. Shao, Y . Mao, and J. Zhang, “Task-oriented communica tion for mul- tidevice cooperative edge inference,” IEEE Trans. Wireless Commun. , vol. 22, no. 1, pp. 73–87, Jan. 2023

    2023

  27. [35]

    MSR-VTT: A large video d escription dataset for bridging video and language,

    J. Xu, T. Mei, T. Y ao, and Y . Rui, “MSR-VTT: A large video d escription dataset for bridging video and language,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2016, pp. 5288–5296

    2016

  28. [36]

    TinyViT: Fast pretraining distillation for small vision t ransformers,

    K. Wu, J. Zhang, H. Peng, M. Liu, B. Xiao, J. Fu, and L. Y uan , “TinyViT: Fast pretraining distillation for small vision t ransformers,” in Proc. Eur . Conf. Comput. Vis. , 2022, pp. 68–85

    2022

  29. [37]

    Qwen3 technical report,

    A. Y ang et al. , “Qwen3 technical report,” arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  30. [38]

    BLEU: A m ethod for automatic evaluation of machine translation,

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “BLEU: A m ethod for automatic evaluation of machine translation,” in Proc. Annu. Meeting Assoc. Comput. Linguistics , 2002, pp. 311–318

    2002

  31. [39]

    CIDEr: Conse nsus-based image description evaluation,

    R. V edantam, C. L. Zitnick, and D. Parikh, “CIDEr: Conse nsus-based image description evaluation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2015, pp. 4566–4575

    2015

  32. [2025]

    Available: https://arxiv.org/abs/2511

    [Online]. Available: https://arxiv.org/abs/2511. 23282

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.