Empowering Intelligent Low-altitude Economy with Large AI Model Deployment

Dong In Kim; Hongyang Du; Jie Xu; Junting Chen; Kaibin Huang; Yulan Gao; Zhonghao Lyu

REVIEW 3 major objections 5 minor 1 cited by

Empowering Intelligent Low-altitude Economy with Large AI Model Deployment

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Prompt-driven large AI models can place drones better than channel-model optimization.

desk verdict A readable systems-architecture synthesis with a real-world drone-channel dataset, but the headline UAV-placement result rests on an unfair baseline comparison. read the letter →

arxiv 2505.22343 v2 pith:RXUVVTFI submitted 2025-05-28 eess.SP cs.AI

Zhonghao Lyu , Yulan Gao , Junting Chen , Hongyang Du , Jie Xu , Kaibin Huang , Dong In Kim This is my paper

classification eess.SPcs.AI

keywords low-altitudeeconomylargeAImodeldeploymentUAVplacementreferencesignalreceivedpowerintegratedsensingcommunicationandcomputationco-inferencereal-worldchannelmeasurement

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 is trying to establish that large AI models (LAIMs) can be deployed to run drone-based services — the so-called low-altitude economy — if they are distributed across ground, aerial, and cloud layers and fed real-world measurements rather than idealized mathematical models. Its strongest evidence is a field experiment in which a prompt-driven LAIM, given measured signal-strength data from a campus testbed, chose UAV positions that beat a conventional optimizer built on an idealized line-of-sight channel model, with the gap growing as more UAVs were added. A second experiment shows that splitting a pruned large model between a drone and the cloud balances inference quality, delay, and energy better than running it entirely in either place. A sympathetic reader would take away that data-driven LAIMs can outperform model-based optimization in real low-altitude environments, and that the practical obstacle is deployment architecture, not the AI itself.

What carries the argument

The mechanism that carries the argument is the prompt-driven interaction loop: the LAIM receives real-world RSRP measurements plus a natural-language task description, interprets the spatial distribution of signal strength, and outputs UAV coordinates with estimated rates. The surrounding framework is a three-tier hierarchical architecture (ground, aerial, and cloud layers) and a task-oriented execution pipeline with offline initialization, multi-modal sensing, information transmission, downstream execution, and closed-loop adaptation, coordinated by integrated sensing, communication, and computation (ISCC). For the second case study, the central object is a pruning-aware co-inference scheme that jointly optimizes the pruning ratio, transmit power, and computation frequency.

What would settle it

Feed the same measured signal-strength map to the conventional optimization baseline as its channel model, and rerun the sum-rate maximization; if the LAIM's advantage disappears or narrows sharply, the claim that LAIMs make better use of real-world measurements than optimization-based placement would be refuted.

Watch

Extended reading notes

Core claim

The paper claims that LAIMs can be integrated into low-altitude systems through a three-tier ground-aerial-cloud architecture and a task-oriented execution pipeline, and that this integration pays off in the field: a prompt-driven LAIM given measured reference-signal-received-power (RSRP) values selected UAV positions whose sum communication rate exceeded that of a successive-convex-approximation optimizer using an idealized line-of-sight channel model, with the gap widening as the number of UAVs increased. It further claims that a pruning-aware co-inference scheme, which partitions a pruned LAIM between a UAV and the cloud, outperforms both fully onboard and fully cloud inference across the trade-off among inference quality, delay, and energy.

Load-bearing premise

The result rests on the claim that the right comparison for a LAIM that receives real measured signal data is an optimization baseline that sees only an idealized line-of-sight channel model and not the same measurements.

Editorial extensions

If this is right

LAIM-guided placement becomes more valuable as fleets grow, because the reported performance gap over the LoS-based optimizer widens with the number of UAVs.
Prompt-driven workflows can replace hand-built optimization routines for placement tasks, since the LAIM reads the task description and the measured map directly.
Pruning-aware co-inference makes it feasible to run large-model intelligence on resource-limited aerial platforms rather than requiring cloud-only execution.
The closed-loop pipeline gives a concrete mechanism for turning lab-trained LAIMs into field-adapted ones using the low-altitude system's own sensing data.

Reading between the lines

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

Not tested in the paper: handing the same measured radio map to the optimization baseline would isolate whether the advantage comes from the data or from the LAIM's interpretation of it.
Prompt design is likely a dominant and under-controlled variable; a differently worded task description could change the LAIM's placements as much as the measurements do.
The same prompt-driven loop plausibly transfers to other physical-layer decisions, such as beam selection, handover, or energy-aware routing, whenever a measured field can be expressed in natural language.

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

The architecture part of this paper is a competent synthesis of ideas already in the literature—ISAC/ICC/ISCC, split inference, continual learning—organized into a three-tier ground/aerial/cloud framework for low-altitude economy. That part is plausible and clearly written. The paper also ships a real-world RSRP measurement campaign (USRP on a DJI M300, 5G SSB beams, RTK positioning), which is genuine effort and gives the case studies a concrete basis. Credit where it is due: the data collection and the prompt-driven UAV placement workflow are not things you see every day in a system paper.

The soft spot is exactly where the stress-test note lands. The Section IV-B comparison gives the LAIM the measured RSRP values while the SCA baseline is handed an idealized LoS channel model. The paper even explains the performance gap as coming from that model mismatch. That is not a fair test of LAIM capability; it is an information-asymmetry setup. A stronger comparison would give the baseline the same measured RSRP (e.g., via a channel knowledge map or interpolation) or at least evaluate both schemes on the same measured channel. Without that, the claim that LAIMs 'significantly outperform' traditional optimization is not established. The paper also reports no error bars or multiple seeds for Figures 4 and 5, which is a minor but relevant omission for a real-world measurement. Section IV-C, the resource-aware co-inference study, lacks a formal optimization formulation, so the reader cannot tell exactly what was optimized or how the three paradigms were compared.

These flaws are real but not fatal to the paper's core contribution, which is architectural. The case studies read as demonstrations of feasibility rather than rigorous empirical proofs. If the authors revise with a fair baseline and explicit problem statements, the framework could be a useful reference for people working on LLM-driven UAV networks. As it stands, I would not cite the empirical result, but I might cite the architecture if I needed a compact overview of the design space.

My recommendation: do not desk-reject. Send it to peer review with a clear request to fix the baseline comparison and report variance. A serious referee could turn this into a solid systems paper, but only if the validation is made honest.

Referee Report

3 major / 5 minor

Summary. This paper proposes a hierarchical three-tier architecture (ground, aerial, cloud) for deploying large AI models (LAIMs) in low-altitude economy (LAE) systems, discusses key enabling techniques (ISAC, ICC, ISCC), presents a task-oriented execution pipeline with offline and online stages, and reports two real-world case studies: LAIM-guided UAV placement using measured RSRP data, and pruning-aware co-inference for resource-limited LAIM deployment. It concludes with six open research directions. The paper is primarily an architecture/vision contribution, with the case studies intended to demonstrate feasibility of the proposed framework.

Significance. If the empirical claims hold, the paper offers a useful architectural synthesis and a concrete demonstration that large AI models can use measured signal-strength maps for environment-aware UAV placement, and that pruning-aware split inference can balance quality, delay, and energy on resource-limited aerial platforms. Strengths include the real-world measurement campaign (USRP-based 5G SSB RSRP collection at 2.6 GHz with RTK localization), the explicit task-oriented pipeline, and the attempt to ground LAIM deployment in measured rather than simulated channels. The significance is tempered by the fact that the central validation in Section IV-B rests on an asymmetric comparison, and Section IV-C does not state its optimization problem; these issues affect the weight that should be placed on the claimed quantitative gains.

major comments (3)

[IV-B] The central empirical claim that "LAIM-based solutions significantly outperform" the SCA baseline rests on an asymmetric comparison. The LAIM receives real-world RSRP measurements, whereas the baseline optimizes a sum-rate objective built on a "location-dependent LoS channel model" that ignores measured shadowing, reflections, and interference. The paper does not state whether the rates in Fig. 4 are computed under the measured RSRP for both schemes or under each scheme's own channel assumption. If the baseline positions are evaluated in the measured channel, the reported gap may be an artifact of giving the LAIM strictly more information; if they are evaluated in the LoS model, the comparison does not answer which method places UAVs better in the real environment. The authors should give the baseline access to the same measured RSRP information (e.g., via a channel knowledge map or interpolated measurements), re-optimize, and then compare; they should also report variance across repeated runs or prompt variations and, where feasible, statistical tests.
[IV-B] The experimental protocol is insufficiently specified for the paper's claimed validation. The prompt-driven workflow is described only in words; the paper does not identify the LAIM used, the prompt template, the number of independent queries, how raw LAIM outputs are parsed and validated as feasible UAV coordinates, or how the "estimated communication rates" are computed. Without these details, the reader cannot assess whether the LAIM's advantage is due to the model's reasoning or to hand-tuned prompt engineering, nor can the result be reproduced.
[IV-C] The resource-aware co-inference study lacks a statement of the optimization problem. The text says the system "jointly optimizes the pruning ratio, transmit power, and computation frequency" under QoS constraints, but no formulation is given: no decision variables, objective functions, constraints, or parameter values. Figure 5 reports that co-inference "consistently achieves the best performance," but no variance or error bars are shown, and the comparison conditions for on-IAA and on-cloud inference (e.g., exact pruning schedules, wireless parameters, BART configuration) are not specified. This section should be either expanded into a self-contained optimization and experimental study or explicitly labeled as a preliminary illustration rather than a validation.

minor comments (5)

[I] Line "corss-country package delivery" contains a typo; it should be "cross-country."
[II, first paragraph] The sentence "LAIM-empowered LAE requires requires a fundamental shift" contains a duplicated word; remove the second "requires."
[Throughout] The abbreviation "UA V" is used inconsistently (sometimes "UAV" or "UAVs"); please standardize to one form and define it at first use.
[IV-A] The term "PBCH-XSS RSRP metric" is unclear; likely "SSB" or "SS/PBCH block" is intended, and the acronym should be defined consistently.
[IV-B] For reproducibility, the authors should consider releasing the processed RSRP radio map or the measurement dataset, or at least describing the data format and access conditions.

Circularity Check

0 steps flagged · score 2.0 of 10

No substantive circularity; the LAIM-vs-SCA comparison in Section IV-B is asymmetric but not equivalent to its inputs by construction.

full rationale

The paper's central contributions are an architecture, an execution pipeline, and two real-world case studies; none of these is a derivation that reduces to its own inputs. In Section IV-B, the LAIM is given measured RSRP values and a task prompt, and it generates UAV coordinates; the SCA baseline optimizes a sum-rate problem under an explicitly stated 'location-dependent line-of-sight (LoS) channel model.' The resulting performance gap is explained by the authors themselves as due to 'the mismatch between the simplified LoS channel models assumed in the traditional optimization-based scheme and the complex real-world wireless environments.' This is a transparent information-asymmetry in the benchmarking, but it is not circular in the technical sense: the LAIM's outputs are not a refit of the baseline's objective, and no fitted parameter is relabeled as a prediction. Section IV-C compares inference paradigms empirically and does not derive its conclusion from its input. The self-citations in the paper, mainly references [6], [8], [12], and [15], support background claims about pre-training/fine-tuning, efficient inference, ISAC, and ISCC; they are not load-bearing for the empirical validation, and there is no imported uniqueness theorem or ansatz. The score of 2 reflects the presence of minor non-load-bearing self-citation rather than any genuine circularity in the derivation chain.

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

No new physical entities are introduced. Free parameters are not explicitly defined; the case-study variables (pruning ratio, transmit power, computation frequency) are optimization variables, but their feasible sets and objective weights are not stated. The supporting axioms are mostly domain assumptions about data availability and the representativeness of the baseline.

assumptions (5)

domain assumption Low-altitude systems continuously generate vast volumes of real-world, multi-modal, spatio-temporal data under diverse physical constraints.
Invoked in Section III-B as the basis for continual LAIM evolution; no quantitative evidence or dataset characterization is provided.
ad hoc to paper A simplified LoS channel model is a representative baseline for traditional optimization-based UAV placement.
Section IV-B uses this model for the SCA benchmark while the LAIM receives real RSRP, making the comparison favorable to the LAIM by construction.
domain assumption LAIMs can autonomously interpret spatial RSRP distributions and produce near-optimal UAV positions from prompts.
Section IV-B assumes the prompt-driven workflow works without a formal algorithm or controlled evaluation; no ablation or sensitivity analysis is provided.
domain assumption Co-inference with pruning can balance inference quality, delay, and energy through joint optimization.
Section IV-C relies on this without stating the optimization problem, constraints, or theoretical guarantees; the claim is supported only by a single figure.
domain assumption The three-tier ground, aerial, and cloud architecture is scalable and adaptive.
Section II-A asserts scalability and adaptability but provides no complexity analysis, capacity planning, or field validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Empowering Intelligent Low-altitude Economy with Large AI Model Deployment." pith.science (2026). https://pith.science/paper/RXUVVTFI

@misc{pith2026250522343,
  author       = {Pith},
  title        = {Pith review of: Empowering Intelligent Low-altitude Economy with Large AI Model Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RXUVVTFI}},
  note         = {Machine review of arXiv:2505.22343}
}

read the original abstract

Low-altitude economy (LAE) represents an emerging economic paradigm that redefines commercial and social aerial activities. Large artificial intelligence models (LAIMs) offer transformative potential to further enhance the intelligence of LAE services. However, deploying LAIMs in LAE poses several challenges, including the significant gap between their computational/storage demands and the limited onboard resources of LAE entities, the mismatch between lab-trained LAIMs and dynamic physical environments, and the inefficiencies of traditional decoupled designs for sensing, communication, and computation. To address these issues, we first propose a hierarchical system architecture tailored for LAIM deployment and present representative LAE application scenarios. Next, we explore key enabling techniques that facilitate the mutual co-evolution of LAIMs and low-altitude systems, and introduce a task-oriented execution pipeline for scalable and adaptive service delivery. Then, the proposed framework is validated through real-world case studies. Finally, we outline open challenges to inspire future research.

Figures

Figures reproduced from arXiv: 2505.22343 by the authors.

**Figure 2.** Execution pipeline for LAIM-empowered LAE services. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Real-world implementation with the channel measurement prototype in low-altitude scenarios. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: LAIM-guided UAV deployment in LAE. C. Resource-aware LAIM Deployment To enable the deployment of LAIMs on resource-limited LAE platforms, we propose a pruning-aware co-inference scheme guided by real-world RSRP measurements. A pretrained LAIM is first pruned and then …

**Figure 5.** Figure 5: Resource-aware LAIM deployment across LAE platforms. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

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. Full citation record

Near-Field Integrated Imaging and Communication in Distributed MIMO Networks
eess.SP 2025-08 unverdicted novelty 5.0 of 10

Distributed MIMO networks can image near-field targets by converting spatial-domain signals to the Fourier domain, using RMA for small objects and sparse Bayesian learning for large scenes.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Embodied AI-empowered low altitude economy: Integrated sensing, communi- cations, computation, and control (ISC3),

Y . Yang, Y . Chen, J. Wang, G. Sun, and D. Niyato, “Embodied AI-empowered low altitude economy: Integrated sensing, communi- cations, computation, and control (ISC3),” 2024. [Online]. Available: https://arxiv.org/pdf/2412.19996

arXiv 2024
[2]

Generative AI based secure wireless sensing for ISAC Networks,

J. Wang, H. Du, Y . Liu, G. Sun, D. Niyato, S. Mao, D. I. Kim, and X. Shen, “Generative AI based secure wireless sensing for ISAC Networks,” IEEE Trans. Inf. Forensics Secur., early access

work page
[3]

Leveraging large language models for integrated satellite-aerial-terrestrial networks: Recent advances and future directions,

S. Javaid, R. Khalil, N. Saeed, B. He, and M. Alouini, “Leveraging large language models for integrated satellite-aerial-terrestrial networks: Recent advances and future directions,” IEEE Open J. Commun. Soc. , vol. 6, pp. 399-432, 2025

work page 2025
[4]

Generative AI-enabled wireless communications for robust low-altitude economy networking,

C. Zhao, J. Wang, R. Zhang, D. Niyato, G. Sun, H. Du, D. Kim, and A. Jamalipour, “Generative AI-enabled wireless communications for robust low-altitude economy networking,” 2025. [Online]. Available: https://arxiv.org/pdf/2502.18118

arXiv 2025
[5]

Cooperative generative AI for UA V-based scenarios: An intelligent cooperative framework,

L. Zhou, W. Feng, Z. Chen, T. Ruan, S. Leng, H. Yang, Y . Fu, and T. Quek, “Cooperative generative AI for UA V-based scenarios: An intelligent cooperative framework,” IEEE Veh. Technol. Mag., early access

work page
[6]

Rethinking resource management in edge learning: A joint pre-training and fine- tuning design paradigm,

Z. Lyu, Y . Li, G. Zhu, J. Xu, H. Vincent Poor, and S. Cui, “Rethinking resource management in edge learning: A joint pre-training and fine- tuning design paradigm,” IEEE Trans. Wireless Commun., vol. 24, no. 2, pp. 1584-1601, Feb. 2025

work page 2025
[7]

Cost-efficient computation offloading in SAGIN: A deep reinforcement learning and perception-aided approach,

Y . Gao, Z. Ye, and H. Yu, “Cost-efficient computation offloading in SAGIN: A deep reinforcement learning and perception-aided approach,” IEEE J. Sel. Areas Commun. , vol. 42, no. 12, pp. 3462-3476, Dec. 2024

work page 2024
[8]

The larger the merrier? Efficient large AI model inference in wireless edge networks,

Z. Lyu, M. Xiao, J. Xu, M. Skoglund, and M. Renzo, “The larger the merrier? Efficient large AI model inference in wireless edge networks,”

work page

Show all 16 references

[9]

Large language model based multi-objective optimization for integrated sensing and communications in UA V networks,

H. Li, M. Xiao, K. Wang, D. I. Kim, and M. Debbah, “Large language model based multi-objective optimization for integrated sensing and communications in UA V networks,”IEEE Wireless Commun. Letters, vol. 14, pp. 979-983, Apr. 2025

2025
[10]

A tutorial on environment-aware communications via channel knowledge map for 6G,

Y . Zeng, J. Chen, J. Xu, D. Wu, X. Xu, S. Jin, X. Gao, D. Gesbert, S. Cui, and R. Zhang, “A tutorial on environment-aware communications via channel knowledge map for 6G,” IEEE Commun. Surv. Tutor , vol. 26, no. 3, pp. 1478-1519, thirdquarter 2024

2024
[11]

A survey on integrated sensing, communication, and computation,

D. Wen, Y . Zhou, X. Li, Y . Shi, K. Huang, and K. B. Letaief, “A survey on integrated sensing, communication, and computation,” IEEE Commun. Surv. Tutor, early access
[12]

Joint maneuver and beamforming design for UA V-enabled integrated sensing and communication,

Z. Lyu, G. Zhu, and J. Xu, “Joint maneuver and beamforming design for UA V-enabled integrated sensing and communication,” IEEE Trans. Wireless Commun., vol. 22, no. 4, pp. 2424-2440, Apr. 2023

2023
[13]

UA V-aided radio map construction exploiting environment semantics,

W. Liu and J. Chen, “UA V-aided radio map construction exploiting environment semantics,” IEEE Trans. Wireless Commun. , vol. 22, no. 9, pp. 6341-6355, Sept. 2023

2023
[14]

Goal-oriented semantic communication with message-sharing NOMA for metaverse,

Y . Cheng, D. Niyato, H. Du, C. Miao, and D. I. Kim, “Goal-oriented semantic communication with message-sharing NOMA for metaverse,” IEEE Wireless Commun., vol. 31, no. 3, pp. 309 - 315, Jun. 2024

2024
[15]

Pushing AI to wireless network edge: An overview on integrated sensing, communication, and computation towards 6G,

G. Zhu, Z. Lyu, X. Jiao, P. Liu, M. Chen, J. Xu, S. Cui, and P. Zhang, “Pushing AI to wireless network edge: An overview on integrated sensing, communication, and computation towards 6G,” Sci. China Inf. Sci. , vol. 66, no. 130301, pp. 1-19, Feb. 2023

2023
[2025]

Available: https://arxiv.org/pdf/2505.09214

[Online]. Available: https://arxiv.org/pdf/2505.09214

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Embodied AI-empowered low altitude economy: Integrated sensing, communi- cations, computation, and control (ISC3),

Y . Yang, Y . Chen, J. Wang, G. Sun, and D. Niyato, “Embodied AI-empowered low altitude economy: Integrated sensing, communi- cations, computation, and control (ISC3),” 2024. [Online]. Available: https://arxiv.org/pdf/2412.19996

arXiv 2024

[2] [2]

Generative AI based secure wireless sensing for ISAC Networks,

J. Wang, H. Du, Y . Liu, G. Sun, D. Niyato, S. Mao, D. I. Kim, and X. Shen, “Generative AI based secure wireless sensing for ISAC Networks,” IEEE Trans. Inf. Forensics Secur., early access

work page

[3] [3]

Leveraging large language models for integrated satellite-aerial-terrestrial networks: Recent advances and future directions,

S. Javaid, R. Khalil, N. Saeed, B. He, and M. Alouini, “Leveraging large language models for integrated satellite-aerial-terrestrial networks: Recent advances and future directions,” IEEE Open J. Commun. Soc. , vol. 6, pp. 399-432, 2025

work page 2025

[4] [4]

Generative AI-enabled wireless communications for robust low-altitude economy networking,

C. Zhao, J. Wang, R. Zhang, D. Niyato, G. Sun, H. Du, D. Kim, and A. Jamalipour, “Generative AI-enabled wireless communications for robust low-altitude economy networking,” 2025. [Online]. Available: https://arxiv.org/pdf/2502.18118

arXiv 2025

[5] [5]

Cooperative generative AI for UA V-based scenarios: An intelligent cooperative framework,

L. Zhou, W. Feng, Z. Chen, T. Ruan, S. Leng, H. Yang, Y . Fu, and T. Quek, “Cooperative generative AI for UA V-based scenarios: An intelligent cooperative framework,” IEEE Veh. Technol. Mag., early access

work page

[6] [6]

Rethinking resource management in edge learning: A joint pre-training and fine- tuning design paradigm,

Z. Lyu, Y . Li, G. Zhu, J. Xu, H. Vincent Poor, and S. Cui, “Rethinking resource management in edge learning: A joint pre-training and fine- tuning design paradigm,” IEEE Trans. Wireless Commun., vol. 24, no. 2, pp. 1584-1601, Feb. 2025

work page 2025

[7] [7]

Cost-efficient computation offloading in SAGIN: A deep reinforcement learning and perception-aided approach,

Y . Gao, Z. Ye, and H. Yu, “Cost-efficient computation offloading in SAGIN: A deep reinforcement learning and perception-aided approach,” IEEE J. Sel. Areas Commun. , vol. 42, no. 12, pp. 3462-3476, Dec. 2024

work page 2024

[8] [8]

The larger the merrier? Efficient large AI model inference in wireless edge networks,

Z. Lyu, M. Xiao, J. Xu, M. Skoglund, and M. Renzo, “The larger the merrier? Efficient large AI model inference in wireless edge networks,”

work page