Pith. sign in

REVIEW 4 major objections 4 minor 32 references

A large language model, prompted with compressed network state, can plan UAV velocities and sensor visits to minimize packet loss; the paper claims this matches a trained multi-agent DQN while halving prompt length and beating a greedy base

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 11:16 UTC pith:WBZMWOXC

load-bearing objection Plausible idea undermined by an unspecified attention training procedure; not ready for peer review. the 4 major comments →

arxiv 2510.05698 v2 pith:WBZMWOXC submitted 2025-10-07 cs.AI

AIC-VDS: Attention-Based In-Context Learning for Joint Velocity Control and Data Collection Scheduling in Multi-UAV-Assisted Pipeline Monitoring

Yousef Emami , Miguel Gutierrez Gaitan , Atefeh Hajijamali Arani , Jingjing Zheng , Hao Zhou This is my paper
classification cs.AI
keywords UAV swarm controlin-context learninglarge language modelsattention mechanismpacket loss minimizationdata collection schedulingvelocity controldisaster monitoring
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 paper tries to establish that a large language model, prompted in natural language with a few examples, can replace online deep reinforcement learning for the joint control of multiple UAVs collecting data from ground sensors in a disaster zone. The key move is to compress the continuous stream of sensor states — queue lengths, battery levels, channel conditions, UAV positions — with an attention module that selects the top-k most urgent sensors before the prompt is built. In simulation, this AIC-VDS scheme matches the packet loss of a trained multi-agent DQN while stabilizing quickly, and beats a greedy maximum-channel-gain policy by a wide margin (91% lower packet loss at 20 sensors), while cutting average prompt length by 50%. A sympathetic reader would care because LLM-based control requires no retraining and no explicit optimization model, which matters when a tsunami or infrastructure emergency leaves no time for DRL training.

Core claim

The paper's central claim is that learned attention scoring is enough to turn a large, noisy, partially outdated sensor-state vector into a short prompt from which an edge-hosted LLM can produce near-optimal UAV actions. Each sensor's features (queue length, battery, channel gain) are projected into query, key, and value vectors; softmax attention produces a context vector; and a linear layer emits a scalar urgency score. The top-k sensors by that score are sent to the LLM, which then outputs a data-collection schedule and a velocity for each UAV, with a feedback loop recording outcomes for future prompts. The authors report that this arrangement minimizes packet loss from both buffer overfl

What carries the argument

The attention-based top-k selector is the load-bearing component. It is a small self-attention module (queries, keys, values with softmax weighting, followed by a linear urgency scorer) that ranks all ground sensors by current urgency and keeps only the k most important for the LLM prompt. This is what lets the system claim both low overhead (50% shorter prompts) and good decisions: the compress-and-then-reason chain is the argument.

Load-bearing premise

The whole system relies on the attention module's top-k selection keeping the sensor information the LLM actually needs; if the learned urgency scores drop a sensor that later overflows or fails, the compressed prompt cannot recover it, and the claimed packet-loss reduction collapses.

What would settle it

Rerun the same simulation with the attention-based selector replaced by a random top-k of identical size, and measure packet loss over many episodes; if random pruning matches AIC-VDS, then the learned attention is not the cause of the gains. Independently, check Eq. (10aa): with the conjunction as written, a communication failure is counted only when the UAV velocity is simultaneously at or below its maximum, so the channel failure term is near-always zero; re-deriving the packet-loss model and re-running the comparison would be a direct test.

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

If this is right

  • If AIC-VDS holds up, emergency UAV fleets can obtain scheduling policies by prompting a pretrained LLM, with no DQN training loop and no task-specific reward engineering.
  • The 50% prompt-length reduction means the bottleneck of LLM input limits is partially relieved by a cheap onboard attention step, making edge-hosted LLM control more plausible in real time.
  • The method extends naturally to more sensors and more UAVs: the reported complexity is quadratic in the pruned sequence length, so the top-k compression keeps LLM calls bounded.
  • If packet loss is truly lower with three UAVs, the framework suggests that small fleets with attention-based ICL can do work that previously needed trained multi-agent RL.
  • Larger proprietary LLMs gave better convergence in the authors' comparison, pointing to model capability as a key determinant of control quality.

Where Pith is reading between the lines

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

  • The 91% headline reduction is measured against the maximum-channel-gain baseline, not against the MADQN baseline; a fair reader should treat the DQN comparison as 'comparable', not 'superior', until more seeds are reported.
  • The same attention-plus-LLM recipe could generalize to other data-ferrying problems — wildfire sensing, smart agriculture, post-earthquake inspection — wherever the state space is large but the decisive features are sparse; the paper does not test this, but nothing in the method ties it to tsunami geometry.
  • A sharper test of the attention module would compare it against a random or round-robin top-k of the same size; the paper only compares with and without attention, so it does not isolate whether the learned scoring, rather than mere pruning, is what helps.
  • Because the paper never specifies the attention loss function, reproducing the urgency scores is not yet possible from the text; an explicit training objective (e.g., supervised on LLM-chosen sensors or on packet-loss regret) would be the missing piece.

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 / 4 minor

Summary. The paper proposes AIC-VDS, an attention-based in-context learning framework for joint velocity control and data-collection scheduling in multi-UAV-assisted sensor networks. The system uses an attention module to compress sensor state information (queue length, battery, channel gain) and selects a top-k subset for an edge-hosted LLM, which then generates schedules and velocities via in-context learning. The authors claim that AIC-VDS outperforms Multi-agent DQN and maximum-channel-gain baselines in packet loss, with a 91% reduction at J=20, and that the attention mechanism halves the prompt length. The central technical components are the attention-based compression and the LLM-driven decision loop, but the paper lacks a specification of how the attention parameters are trained, among other issues.

Significance. If the claimed results were fully supported, the paper would offer a potentially useful alternative to DRL for time-critical UAV data collection, with the practical advantage of avoiding online training. The attention-based input reduction addresses a genuine limitation of LLM-based controllers. However, the contribution is currently not verifiable: the attention training objective is undefined, the packet-loss model contains a likely erroneous condition, and the empirical evaluation lacks statistical rigor. The paper also has a direct inconsistency between the arXiv abstract (pipeline monitoring) and the full text (tsunami/post-disaster monitoring). As such, the significance cannot be assessed from the manuscript in its present form.

major comments (4)
  1. [Algorithm 1, step 5 (Sec. V-B)] The attention mechanism is a central component, with learnable matrices W_Q, W_K, W_V (Eq. 13) and w_s, b_s (Eq. 17). However, Algorithm 1 step 5 only says 'Compute attention probabilities and update parameters' without specifying a loss function, training data, or optimizer. Without this, the attention weights are undefined and the 91% packet-loss improvement in Fig. 6c cannot be reproduced or attributed to the described mechanism. The comparison against the no-attention variant in Fig. 6a shows at most that some form of input reduction helps, not that this attention module is responsible.
  2. [Sec. IV-A, Eq. (10aa)] The communication-failure term f_i^j(t,v) is defined as 1 when (t = t_i^j) AND (gamma_i^j <= gamma_th) AND (v_i <= v_i^max). Since v_i^max is the maximum velocity, the condition v_i <= v_i^max is always satisfied, making it vacuously true. If the intent is that low velocity causes data loss, that contradicts the introduction's claim that slow UAV movement exacerbates buffer overflow; if the intent is that high velocity causes communication errors, the inequality should be reversed. As written, this term degenerates and the packet-loss model is not credible.
  3. [Sec. VI-C, Fig. 6] All simulation results are presented without error bars, confidence intervals, or any indication of the number of random seeds or independent runs. The text claims a 91% lower packet loss for AIC-VDS with I=3 versus maximum channel gain at J=20, but the corresponding figure caption (Fig. 6c) says 'Packet loss across individual ground sensors,' while the body text says 'network costs as a function of the number of ground sensors.' This ambiguity, together with the lack of statistical evidence, makes the central empirical claim non-reproducible and potentially overstated.
  4. [Title/Abstract vs. Full Text] The arXiv title and abstract emphasize 'pipeline monitoring,' while the full-text title and introduction describe 'post-disaster monitoring' and tsunami scenarios. The conclusion mentions 'smoke density and imagery' as additional sensory data, but the system model and simulation use only queue length, battery, and channel gain. These contradictions confuse the application context and the actual scope of the contribution.
minor comments (4)
  1. [Sec. V-C, Eqs. (20)-(21)] The complexity expression is unclear: O(m^2 + d·m) + O(n^2 + k·n) is followed by 'thus, the final complexity is O(I·(n^2 + k·n)).' The relationship among m, n, d, k, and I is not defined, and the 'therefore' is not obvious. Please clarify the derivation and the meaning of each symbol.
  2. [Throughout] Several typographical issues: 'In-Contetxt Learning' in the Section III heading, inconsistent spacing in 'UA V', and the repeated phrase 'the exponential application in urgent or rapidly changing situations' in Sec. II-A. The manuscript needs careful proofreading.
  3. [Sec. VI-B, Baselines] For the MADQN baseline, the text states only that 'the performance of the MADQN is given for the last episode.' This is insufficient to understand whether MADQN was trained to convergence, how many episodes, or what hyperparameters were used. A fair comparison requires the same evaluation protocol for all methods.
  4. [Sec. I, Contribution bullet] The third contribution bullet states a specific numerical result (91% lower packet loss) in the introduction. This is unusual for a contribution list and should be moved to the results section; also, as written, it depends on the unverified attention training process.

Circularity Check

0 steps flagged

No significant circularity: AIC-VDS is evaluated empirically; the attention-module training omission is a reproducibility concern, not a circular reduction.

full rationale

After walking the derivation chain, I find no instance where a predicted quantity or first-principles result reduces to an input by construction. The optimization objective (Eq. 10a) is a defined cost function, not a derived prediction; the attention mechanism (Eqs. 11-19) is a stated architecture for top-k selection, and the 50% prompt-length reduction and 91% packet-loss improvement are empirical measurements reported in Section VI, not analytical consequences of fitted parameters. Algorithm 1 step 5 ('Compute attention probabilities and update parameters') does omit the training loss/optimizer, which makes the attention module underspecified and hinders reproducibility, and the paper's related work cites several prior works by the same authors (Refs. [3], [16]-[19]); however these are contextual/incremental citations and the MADQN/MCG baselines are independently implemented comparisons, so no load-bearing claim reduces to a self-citation. The manuscript itself acknowledges at the end that real-world robustness requires further work. Because no equation is equivalent to another by construction and no fitted parameter is renamed as a prediction, the appropriate score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claim depends on unstated training of the attention module, an untested compression-fidelity assumption, and a simplified simulation model. The free parameters and ad-hoc assumptions are concentrated in the attention/LLM interface.

free parameters (3)
  • Attention weight matrices W_Q, W_K, W_V = unknown
    Eq. 13 projection weights are learnable and updated in Algorithm 1 step 5, but no loss function or training data is specified.
  • Importance linear layer w_s, b_s = unknown
    Eq. 17 parameters that map attention context to a scalar importance score; also trained without a specified objective.
  • Top-k value k = not specified
    Number of sensors retained for the LLM prompt; directly controls the 50% prompt reduction claim, but no value or sensitivity analysis is given.
axioms (6)
  • domain assumption The LoS probability model of Eq. 7 (with constants a,b) accurately represents air-to-ground channel conditions.
    Borrowed from Al-Hourani et al. [30]; no validation in this scenario.
  • domain assumption The path loss model of Eq. 9 (from Emami et al. [31]) captures packet-loss causes.
    Self-cited model; used to compute gamma in the objective.
  • ad hoc to paper An LLM, given compressed natural-language prompts, can generate near-optimal data collection schedules and velocities.
    Core assumption of the method; no theoretical or systematic empirical support beyond the reported simulations.
  • ad hoc to paper The top-k attention representation preserves sufficient information for the LLM's decisions.
    The paper asserts this but provides no ablation or information-theoretic justification.
  • domain assumption Packet loss is fully captured by the binary functions f and g in Eq. (10aa)-(10ab).
    Simplification; ignores retransmissions, interference, and continuous loss probabilities.
  • ad hoc to paper The attention module can be trained (Algorithm 1 step 5) without a defined loss function.
    The update rule for attention probabilities is never specified.

pith-pipeline@v1.3.0-alltime-deepseek · 12389 in / 12745 out tokens · 105489 ms · 2026-08-04T11:16:28.957015+00:00 · methodology

0 comments
read the original abstract

Uncrewed aerial vehicles (UAVs) are increasingly deployed for autonomous inspection and sensor data collection in large-scale infrastructure monitoring applications, such as pipeline monitoring, where timely anomaly detection is critical. Jointly optimizing data-collection schedules and flight velocities is a critical challenge, as inefficiencies can increase packet loss and inspection latency. While online deep reinforcement learning (DRL) is a widely investigated approach, it suffers from low sample efficiency, substantial training requirements, and simulation-to-reality gaps in time-sensitive scenarios. Large language models (LLMs) offer a promising alternative through in-context learning (ICL); however, their substantial input requirements can introduce considerable computational and communication overhead. To address this, we propose Attention-Based In-Context Learning for Velocity Control and Data Collection Scheduling (AIC-VDS), a joint optimization framework designed to minimize packet loss under partial and potentially outdated local network-state information. AIC-VDS utilizes an attention module to process real-time network-state data, including sensor battery levels, sensor queue lengths, communication channel conditions, UAV locations, time since the previous sensor visit, and sensor urgency scores. This module extracts task-relevant features to reduce input overhead before querying the LLM. The LLM leverages these compressed natural-language prompts to generate adaptive data-collection schedules and velocity-control decisions for UAV execution. Simulation results show that the attention-based representation reduces the average prompt length by 50\%, while AIC-VDS rapidly stabilizes packet loss in the considered scenario.

Figures

Figures reproduced from arXiv: 2510.05698 by Atefeh Hajijamali Arani, Hao Zhou, Jingjing Zheng, Miguel Gutierrez Gaitan, Yousef Emami.

Figure 1
Figure 1. Figure 1: A multi-UAV-assisted tsunami monitoring. Each UAV [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall structure of the proposed AIC-VDS. The MUPDM interacts with an edge-hosted LLM via structured prompts. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Communication protocol for the proposed AIC-VDS. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Flowchart of the AIC-VDS algorithm, illustrating the iterative process for generating optimized data collection schedules [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: C. Complexity The overall complexity of the proposed AIC-VDS is O [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attention mechanism for sensor ranking. Sensor features [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Experimental evaluation of the proposed AIC-VDS framework. (a) (b) Convergence analysis. (c) (d) Performance [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Velocities of AIC-VDS I=5 (a) UAV 1 (b) UAV 2 (c) UAV 3 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparative analysis of LLM inference latency across three UAVs on a circular trajectory. [PITH_FULL_IMAGE:figures/full_fig_p010_8.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

32 extracted references · 7 linked inside Pith

  1. [1]

    Multi uav based traffic control in smart cities,

    B. Y . Yilmaz and S. N. Denizer, “Multi uav based traffic control in smart cities,” in11th International Conference on Computing, Communication and Networking Technologies (ICCCNT), Kharagpur, India, Jul. 2020, pp. 1–7

    2020

  2. [2]

    Unmanned aerial vehicles for package delivery and network coverage,

    M. Khosravi and H. Pishro-Nik, “Unmanned aerial vehicles for package delivery and network coverage,” inIEEE 91st V ehicular Technology Conference (VTC2020-Spring), Antwerp, Belgium, May 2020, pp. 1–5

    2020

  3. [4]

    Topographic data acquisition in tsunami-prone coastal area using unmanned aerial vehicle (uav),

    M. Marfai, N. Khakim, A. Cahyadi, F. Rosaji, H. Fatchurohman, Y . Wi- bowoet al., “Topographic data acquisition in tsunami-prone coastal area using unmanned aerial vehicle (uav),” inIOP Conference Series: Earth and Environmental Science, vol. 148, no. 1. IOP Publishing, 2018, p. 012004

    2018

  4. [5]

    (2025) Tsunamis – health impacts and who response

    World Health Organization. (2025) Tsunamis – health impacts and who response. Accessed: 2025-09-14. [Online]. Available: https: //www.who.int/health-topics/tsunamis#tab=tab 1

    2025

  5. [6]

    Generalising rescue operations in disaster scenarios using drones: A lifelong reinforcement learning approach,

    J. Xu, D. Panagopoulos, A. Perrusqu ´ıa, W. Guo, and A. Tsourdos, “Generalising rescue operations in disaster scenarios using drones: A lifelong reinforcement learning approach,”Drones, vol. 9, no. 6, p. 409, 2025

    2025

  6. [7]

    Deep reinforcement learning: A survey,

    X. Wang, S. Wang, X. Liang, D. Zhao, J. Huang, X. Xu, B. Dai, and Q. Miao, “Deep reinforcement learning: A survey,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 4, pp. 5064–5078, Sep. 2024

    2024

  7. [8]

    Data-driven flight control of internet-of-drones for sensor data aggregation using multi-agent deep reinforcement learning,

    K. Li, W. Ni, Y . Emami, and F. Dressler, “Data-driven flight control of internet-of-drones for sensor data aggregation using multi-agent deep reinforcement learning,”IEEE Wireless Communications, vol. 29, no. 4, pp. 18–23, Oct. 2022

    2022

  8. [9]

    A review on large language models: Architectures, applications, taxonomies, open issues and challenges,

    M. A. K. Raiaan, M. S. H. Mukta, K. Fatema, N. M. Fahad, S. Sakib, M. M. J. Mim, J. Ahmad, M. E. Ali, and S. Azam, “A review on large language models: Architectures, applications, taxonomies, open issues and challenges,”IEEE Access, vol. 12, pp. 26 839–26 874, Feb. 2024

    2024

  9. [10]

    Large language model (llm)-enabled in-context learning for wireless network optimization: A case study of power control,

    H. Zhou, C. Hu, D. Yuan, Y . Yuan, D. Wu, X. Liu, and C. Zhang, “Large language model (llm)-enabled in-context learning for wireless network optimization: A case study of power control,”arXiv preprint arXiv:2408.00214, 2024

    Pith/arXiv arXiv 2024

  10. [11]

    Deep reinforcement learning for joint cruise control and in- telligent data acquisition in uavs-assisted sensor networks,

    Y . Emami, “Deep reinforcement learning for joint cruise control and in- telligent data acquisition in uavs-assisted sensor networks,”arXiv preprint arXiv:2312.09953, 2023

    Pith/arXiv arXiv 2023

  11. [12]

    Online velocity control and data capture of drones for the internet of things: An onboard deep reinforcement learning approach,

    K. Li, W. Ni, E. Tovard, and A. Jamalipour, “Online velocity control and data capture of drones for the internet of things: An onboard deep reinforcement learning approach,”IEEE V ehicular Technology Magazine, vol. 16, no. 1, pp. 49–56, Dec. 2021

    2021

  12. [13]

    The mystery of in- context learning: A comprehensive survey on interpretation and analysis,

    Y . Zhou, J. Li, Y . Xiang, H. Yan, L. Gui, and Y . He, “The mystery of in- context learning: A comprehensive survey on interpretation and analysis,” 2024

    2024

  13. [14]

    Llm-based edge intelligence: A comprehensive sur- vey on architectures, applications, security and trustworthiness,

    O. Friha, M. Amine Ferrag, B. Kantarci, B. Cakmak, A. Ozgun, and N. Ghoualmi-Zine, “Llm-based edge intelligence: A comprehensive sur- vey on architectures, applications, security and trustworthiness,”IEEE Open Journal of the Communications Society, vol. 5, pp. 5799–5856, Sep. 2024

    2024

  14. [15]

    Aero-llm: A distributed framework for secure uav communication and intelligent decision-making,

    B. Dharmalingam, R. Mukherjee, B. Piggott, G. Feng, and A. Liu, “Aero-llm: A distributed framework for secure uav communication and intelligent decision-making,” 2025

    2025

  15. [16]

    Llm-enabled in-context learning for data collection scheduling in uav-assisted sensor networks,

    Y . Emami, H. Zhou, S. Nabavirazani, and L. Almeida, “Llm-enabled in-context learning for data collection scheduling in uav-assisted sensor networks,”arXiv preprint arXiv:2504.14556, 2025

    arXiv 2025

  16. [17]

    From prompts to protection: Large language model-enabled in-context learning for smart public safety uav,

    Y . Emami, H. Zhou, M. G. Gaitan, K. Li, L. Almeida, and Z. Han, “From prompts to protection: Large language model-enabled in-context learning for smart public safety uav,”arXiv preprint arXiv:2506.02649, 2025

    arXiv 2025

  17. [18]

    Frsicl: Llm-enabled in-context learning flight resource allocation for fresh data collection in uav-assisted wildfire monitoring,

    Y . Emami, H. Zhou, M. G. Gaitan, K. Li, and L. Almeida, “Frsicl: Llm-enabled in-context learning flight resource allocation for fresh data collection in uav-assisted wildfire monitoring,”arXiv preprint arXiv:2507.10134, 2025

    arXiv 2025

  18. [19]

    Large language models (llms) for wireless networks: An overview from the prompt engineering perspective,

    H. Zhou, C. Hu, D. Yuan, Y . Yuan, D. Wu, X. Chen, H. Tabassum, and X. Liu, “Large language models (llms) for wireless networks: An overview from the prompt engineering perspective,”arXiv preprint arXiv:2411.04136, 2024

    Pith/arXiv arXiv 2024

  19. [20]

    A survey on in-context learning,

    Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, T. Liuet al., “A survey on in-context learning,”arXiv preprint arXiv:2301.00234, 2022

    Pith/arXiv arXiv 2022

  20. [21]

    Large language models in wireless application design: In-context learning-enhanced au- tomatic network intrusion detection,

    H. Zhang, A. B. Sediq, A. Afana, and M. Erol-Kantarci, “Large language models in wireless application design: In-context learning-enhanced au- tomatic network intrusion detection,” 2024

    2024

  21. [22]

    Leveraging large language models for wireless symbol detection via in-context learning,

    M. Abbas, K. Kar, and T. Chen, “Leveraging large language models for wireless symbol detection via in-context learning,” 2024

    2024

  22. [23]

    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

  23. [24]

    Generative ai for unmanned vehicle swarms: Challenges, applications and opportunities,

    G. Liu, N. Van Huynh, H. Du, D. T. Hoang, D. Niyato, K. Zhu, J. Kang, Z. Xiong, A. Jamalipour, and D. I. Kim, “Generative ai for unmanned vehicle swarms: Challenges, applications and opportunities,” arXiv preprint arXiv:2402.18062, 2024

    Pith/arXiv arXiv 2024

  24. [25]

    Uavs meet llms: Overviews and perspectives toward agentic low-altitude mobility,

    Y . Tian, F. Lin, Y . Li, T. Zhang, Q. Zhang, X. Fu, J. Huang, X. Dai, Y . Wang, C. Tianet al., “Uavs meet llms: Overviews and perspectives toward agentic low-altitude mobility,”arXiv preprint arXiv:2501.02341, 2025

    Pith/arXiv arXiv 2025

  25. [26]

    Large language models for uavs: Current state and pathways to the future,

    S. Javaid, H. Fahim, B. He, and N. Saeed, “Large language models for uavs: Current state and pathways to the future,”IEEE Open Journal of V ehicular Technology, vol. 5, pp. 1166–1192, Aug. 2024

    2024

  26. [27]

    Net-gpt: A llm-empowered man-in-the-middle chatbot for unmanned aerial vehicle,

    B. Piggott, S. Patil, G. Feng, I. Odat, R. Mukherjee, B. Dharmalingam, and A. Liu, “Net-gpt: A llm-empowered man-in-the-middle chatbot for unmanned aerial vehicle,” inProceedings of the Eighth ACM/IEEE Symposium on Edge Computing, Wilmington, DE, USA, Feb. 2023, pp. 287–293

    2023

  27. [28]

    Enhancing autonomous system security and resilience with generative ai: A compre- hensive survey,

    M. Andreoni, W. T. Lunardi, G. Lawton, and S. Thakkar, “Enhancing autonomous system security and resilience with generative ai: A compre- hensive survey,”IEEE Access, vol. 12, pp. 109 470–109 493, Aug. 2024

    2024

  28. [29]

    Large model based agents: State-of-the-art, cooperation paradigms, security and privacy, and future trends,

    Y . Wang, Y . Pan, Z. Su, Y . Deng, Q. Zhao, L. Du, T. H. Luan, J. Kang, and D. Niyato, “Large model based agents: State-of-the-art, cooperation paradigms, security and privacy, and future trends,” 2025

    2025

  29. [30]

    Optimal lap altitude for maximum coverage,

    A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal lap altitude for maximum coverage,”IEEE wireless communications letters, vol. 3, no. 6, pp. 569–572, Jul. 2014

    2014

  30. [31]

    Joint communication scheduling and velocity control in multi-uav-assisted sensor networks: A deep reinforcement learning approach,

    Y . Emami, B. Wei, K. Li, W. Ni, and E. Tovar, “Joint communication scheduling and velocity control in multi-uav-assisted sensor networks: A deep reinforcement learning approach,”IEEE Transactions on V ehicular Technology, vol. 70, no. 10, pp. 10 986–10 998, Sep. 2021

    2021

  31. [32]

    On-board deep q-network for uav-assisted online power transfer and data collection,

    K. Li, W. Ni, E. Tovar, and A. Jamalipour, “On-board deep q-network for uav-assisted online power transfer and data collection,”IEEE Trans- actions on V ehicular Technology, vol. 68, no. 12, pp. 12 215–12 226, Oct. 2019

    2019

  32. [33]

    Joint flight cruise control and data collection in uav-aided internet of things: An onboard deep reinforcement learning approach,

    K. Li, W. Ni, E. Tovar, and M. Guizani, “Joint flight cruise control and data collection in uav-aided internet of things: An onboard deep reinforcement learning approach,”IEEE Internet of Things Journal, vol. 8, no. 12, pp. 9787–9799, Aug. 2021

    2021