Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Hierarchical and Collaborative LLM-Based Control for Multi-UAV Motion and Communication in Integrated Terrestrial and Non-Terrestrial Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that a two-level LLM controller, with a HAPS meta-agent for base-station association and per-UAV agents for motion and telecom actions, beats deep-RL baselines in a simulated 3D aerial highway by achieving higher…

desk verdict The hierarchical LLM architecture is new and the simulations look plausible, but the paper never ablates against a rule-based controller, so the central claim that LLM knowledge drives the gains is unsupported. read the letter →

arxiv 2506.06532 v1 pith:GFXT43R6 submitted 2025-06-06 cs.LG cs.AIcs.NIcs.ROcs.SYeess.SY

classification cs.LGcs.AIcs.NIcs.ROcs.SYeess.SY
keywords hierarchicalLLMcontrolmulti-UAVsystemsaerialhighwaynon-terrestrialnetworksHAPSjointmotionandcommunicationin-contextlearningcollisionavoidance
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

This paper asks whether a pair of pre-trained large language models, wired together in a hierarchy, can replace task-specific deep reinforcement learning for the joint control of UAV motion and communication in a simulated 3D aerial highway. A meta-controller LLM on a high-altitude platform station (HAPS) decides which UAVs connect to which base stations, while an onboard LLM on every UAV chooses acceleration, lane changes, and base-station selection policies. The authors report that this purely prompt-based, no-fine-tuning system yields higher cumulative rewards, lower transportation and telecommunication costs, and substantially lower collision rates than several deep-RL baselines as UAV density scales from 5 to 40. If the result holds, it means general-purpose language models can serve as drop-in controllers for tightly coupled multi-agent mobility and connectivity problems, adapting to new operational rules via prompt edits instead of retraining.

What carries the argument

The load-bearing object is the hierarchical collaborative LLM control architecture. At the top, a HAPS-resident meta-controller LLM treats UAV association as a Markov decision process: its state is HAPS load, per-UAV data rates, and ground coverage; its action is a set of UAVs to offload or recall; its reward penalizes HAPS saturation and handovers while rewarding weighted throughput. At the bottom, each UAV embeds an LLM whose action space is the product of five transport actions and three telecom policies; its reward combines a normalized speed term, a collision penalty, a lane-change frequency penalty, and a handover-aware weighted data rate. The mechanism that carries the argument is in-context learning: every decision is prompted with a task description, a discretized state matrix of surrounding UAVs, and a short list of similar good/bad experiences retrieved from a replay buffer by Euclidean distance, with no gradient updates. The structured prompts convert the optimization problem into a text-completion task.

What would settle it

Run the same simulator with the LLM agents replaced by a deterministic rule-based controller that implements exactly the heuristics in Algorithm 1 and the prompt rules: offload the UAV with the lowest weighted rate when HAPS bandwidth usage exceeds capacity, recall eligible UAVs when capacity frees, choose the fastest action when the ego UAV is centered and has a safe gap, and pick the telecom policy by the same load-threshold logic. If this controller matches or surpasses the LLM-LLM agent's total reward, cost, and collision rate across all UAV densities, then the paper's claim that LLM reasoning drives the improvement is falsified; a complementary check is to compare the LLM's chosen actions to these rules on a held-out set of states and see whether they are nearly identical.

Watch

Extended reading notes

Core claim

The central claim is that a hierarchical pair of collaborative LLM agents—a HAPS-level meta-controller performing UAV-to-BS association control and per-UAV LLMs performing motion and telecom actions—outperforms established deep reinforcement learning baselines on a joint motion-communication control task in a simulated aerial highway. In the proposed architecture, the HAPS LLM continuously monitors backhaul usage; when aggregate demand approaches capacity, it offloads the UAVs with the poorest HAPS link quality to terrestrial base stations, and recalls them when bandwidth frees. Each UAV LLM then chooses one of five transport actions (left-lane change, keep lane, right-lane change, accelerate, decelerate) and one of three telecom policies (best weighted-rate BS, next-candidate-on-saturation, highest instantaneous rate) based on a discretized state matrix and curated good/bad experience examples. The experiments report that the LLM-LLM dual agent converges about an order of magnitude faster than a deep Q-network baseline, reaches a total reward of 30 versus 23 for that baseline and below 20 for a multi-objective RL baseline, and keeps collision rates below 0.08 at high density where the deep Q-network baseline exceeds threshold.

Load-bearing premise

The central claim assumes that the LLM's pretrained knowledge and in-context reasoning produce the performance gains, but the prompts already embed explicit decision rules and curated examples, and the paper does not ablate against a rule-based controller that follows the same heuristics.

Editorial extensions

If this is right

  • If the central claim is correct, a purely prompt-based LLM system can outperform trained deep-RL agents on a joint motion-communication task in a simulated aerial highway, with no task-specific fine-tuning.
  • The hierarchical separation of a global HAPS association controller and local UAV motion/telecom controllers is sufficient to keep collision rates below 0.08 even when UAV density reaches 40.
  • The LLM-LLM approach converges within roughly $1.5\times 10^3$ episodes, about an order of magnitude faster than the DRL baselines, so operators could deploy it with far less simulated training time.
  • Because the system adapts through prompts and in-context examples, it can absorb new safety constraints or operational rules by editing the prompt text rather than retraining a policy network.

Reading between the lines

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

  • A fair test of the causal role of the LLM would replace it with a deterministic controller that executes the heuristics written in the prompts (e.g., offload the lowest-rate UAV at capacity); if that controller matches the reported rewards, the paper's claim that pretrained knowledge drives the improvement would need to be withdrawn, because the prompts already encode the policy.
  • The experimental setup compares against RL agents that are not given the hand-crafted association heuristics, so the comparison may be measuring prompt engineering rather than general intelligence; a stronger baseline would be a DRL agent trained with the same reward shaping and action masks.
  • The paper implies the framework can adapt to novel scenarios via prompt edits, but it does not test this; a concrete extension would vary the highway geometry, channel statistics, or action set and measure whether the same prompts still dominate baselines without any prompt revision.
  • The retrieved experience examples are selected by Euclidean distance in the raw state space; using a learned state embedding could improve retrieval quality and is a natural next step, but it would also introduce training, weakening the no-training claim.
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

4 major / 5 minor

Summary. The paper proposes a hierarchical two-level LLM-based control framework for multi-UAV systems in an aerial highway scenario integrated with terrestrial BSs and a HAPS. A HAPS-level LLM meta-controller performs UAV-to-base-station association and offload/recall decisions, while per-UAV onboard LLMs select motion actions (accelerate, decelerate, lane change, idle) and telecommunication policies (t1/t2/t3). The system model includes 3GPP-style G2A channels, a UAV-HAPS LoS channel model, and a weighted data-rate metric with handover penalties. Simulation results are reported against DDQN, Envelope-MORL, and LLM+DDQN baselines, claiming higher total reward, lower transportation and communication costs, and reduced collision rates. The paper also includes appendices with the exact prompts used for both LLM agents.

Significance. If the central claim were established, the paper would contribute a concrete demonstration that pure prompting-based LLM agents, without fine-tuning, can outperform DRL agents on a joint motion-communication control task in a simulated aerial highway. The manuscript is transparent about its system model and, importantly, releases the full prompt templates in the appendix, which is a strength for reproducibility of the prompting protocol. The hardware/software setup (Ollama, LLM model names, GPU types, episode length) is also reported. However, the experiments lack error bars, random-seed counts, statistical tests, and, most critically, an ablation separating the LLM's pretrained reasoning from the deterministic heuristics already embedded in the prompts and in Algorithm 1. As presented, the evidence does not isolate the contribution of LLM knowledge, and the 'training-phase convergence' curves are not interpretable as learning curves because no learning mechanism is specified for the LLM agents.

major comments (4)
  1. [Section 4.2, Figs. 3(a)-(d)] The 'training-phase convergence' curves are interpreted as evidence of learning ('efficient exploration and stable decision-making early in training'), but no learning mechanism is specified for the LLM agents. Algorithm 1 ends with 'Update policy π_meta using (s_t,a_t,r_t,s_{t+1})' and Section 3.2 refers to MDPs, yet the LLMs are fixed inference models run through Ollama with in-context prompting; there is no fine-tuning, no parameter update, and no defined in-context learning loop that would produce convergence of a policy. The convergence interpretation is therefore unsupported.
  2. [Appendix A and Algorithm 1] The prompts already encode the decision rules that the paper attributes to pretrained LLM reasoning. The HAPS meta-controller is specified as offloading the UAV with the lowest data rate/priority when B_t>C and recalling when spare capacity exists, and the example LLM response in Appendix A follows exactly this rule; the edge-UAV prompt similarly fixes the action set, supplies good/bad experience examples, and includes rule-like guidance. Since there is no ablation against a deterministic rule-based controller that executes the same heuristics (including the nearest-neighbor experience lookup described in Appendix B), the experiments cannot distinguish the contribution of LLM knowledge from the hand-crafted prompt heuristics. The abstract's claim that the method 'achieves ... significantly reduced UAV collision rates' is not attributable to the LLM component on the current evidence.
  3. [Sections 4.2 and 4.3] No error bars, confidence intervals, number of random seeds, or significance tests are reported for any figure. The abstract's word 'significantly' and the conclusion's 'clear Pareto improvement' require statistical support, especially because the reported differences in Fig. 3(d) and Fig. 4(b) appear small relative to the visible baseline variability. Without this information, the central empirical claim is not established beyond a single-run comparison.
  4. [Section 4.1, Fig. 4] The baselines are not sufficiently specified for a fair comparison. DDQN and Envelope-MORL are cited to prior work, but no hyperparameters, network architectures, reward-weight settings, or training budgets are provided, and the 'LLM+DDQN' baseline appears in Fig. 4 as 'Llama 3.1 8B + DDQN' and 'Llama 3.1 70B + DDQN' without explaining how the hybrid is constructed relative to the description in Section 4.1. The reported 16.3% average improvement is therefore difficult to evaluate for baseline fairness.
minor comments (5)
  1. [Section 2.1, Eq. (3) description] The sentence introducing Eq. (3) reads 'The, the gain of antenna element is' and should be rewritten as 'Then, the antenna element gain is'.
  2. [Sections 4.1 and 4.3] The baseline name is inconsistent: 'Envelope-MORL' in Section 4.1 and 'Envelope-MHRL' in the text describing Figs. 3 and 4; the abbreviation should be unified.
  3. [Appendix A] Appendix A labels the example input as 'ChatGPT-4o', while the experiments use the Ollama framework with Llama 3.1 models; clarify which model actually generated the reported results and whether the appendix is illustrative.
  4. [References] The reference listing for Cherif's PhD thesis is incomplete (missing institution formatting and page details), and several references have inconsistent capitalization; a careful copyedit is needed.
  5. [Section 4.3(d)] The claim that the dual-agent system 'requires fewer additional steps compared to DDQN' is not clearly evidenced by the description of Fig. 4(d); clarify whether higher step count is good (longer collision-free trajectories) or bad (more queuing delay), and state the direction of the benefit explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper makes an empirical comparison, not a derivation that reduces to fitted inputs or self-citation.

full rationale

The paper's central claim is an empirical performance comparison between hierarchical LLM control and DRL baselines, not a claimed derivation from first principles. The LLM prompts encode explicit heuristics (Algorithm 1: if B_t > C then offload the UAV with the lowest data rate, else reattach; Appendix A example 'Offload{4}'), but the paper never derives the reported rewards from these heuristics by construction, nor does it fit a parameter and then rename that fit as a prediction. The absence of an ablation against a rule-based controller following the same heuristics weakens the attribution of gains to pretrained LLM knowledge, and the unspecified 'Update policy π_meta' in Algorithm 1 is an implementation/completeness gap, but both are external-validity or correctness concerns, not circularity. Self-citations (e.g., Yan et al. 2023 for the kinematic model, Yan & Tabassum baselines) are not load-bearing for the core claim. Under the required standard of exhibiting a specific reduction (Eq. X = Eq. Y by construction or a fitted parameter renamed as prediction), no circular step can be identified, so the score is 0.

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

The ledger is dominated by hand-tuned reward weights and prompt parameters. The model does not introduce new physical entities. The most structurally important assumptions are that a generic LLM can execute the hand-crafted heuristics well and that the pretrained knowledge adds value beyond those heuristics.

free parameters (6)
  • Reward weights w1-w4
    Hand-chosen weights in transportation and telecommunication reward functions (Eq. 14-15) that balance safety, speed, and connectivity; exact values are not reported in the paper.
  • Meta-reward coefficients eta1-eta3
    Weights in the HAPS meta-controller reward (Eq. 13) controlling the trade-off among throughput, saturation penalty, and handover cost; not reported numerically.
  • Handover penalty coefficient mu = 0 or 0.5
    Mu is set to 0 when keeping the previous BS and 0.5 when switching between terrestrial BS and HAPS; this choice is part of the reward shaping and affects the weighted data rate in Eq. 12.
  • HAPS capacity threshold C = 100 Mbps
    The prompt specifies a total HAPS load limit of 100 Mbps, used to trigger offload decisions; this value is arbitrary and scenario-specific.
  • Number of in-context experience examples = 5
    The agents supply only the five most similar good and bad experiences, determined by Euclidean distance, to limit compute; this number is chosen without sensitivity analysis.
  • Episode length = 30 time steps
    The maximum duration of each simulation episode is capped at 30 time steps, chosen without justification in the paper.
assumptions (6)
  • domain assumption LLM pretrained knowledge generalizes to novel UAV control states through in-context learning
    The paper assumes that a generic LLM can infer reasonable acceleration, lane-change, and telecom actions purely from prompt instructions and a few examples, without task-specific fine-tuning. No experiments isolate whether this holds beyond the specific scripted heuristics.
  • domain assumption The 3GPP TR 36.777 antenna pattern and LoS probability model accurately represent UAV-to-ground propagation
    Section 2.1 adopts the 3GPP model for BS 3D radiation and LoS probability, but does not validate it against measurements for the considered altitudes and urban scenario.
  • domain assumption The Intelligent Driver Model (IDM) adequately characterizes multi-UAV aerial traffic dynamics
    Section 2 uses the continuous IDM from Treiber and Kesting to model acceleration, deceleration, and lane changes. This model is originally designed for road vehicles; its suitability for UAV dynamics is not discussed.
  • ad hoc to paper The hand-crafted prompt rules are near-optimal for the considered scenario
    Algorithm 1 hard-codes the offload/recall decisions as simple threshold-based rules, and the prompt examples are manually selected. The paper implicitly assumes these rules are good enough to produce superior performance, but never compares against a pure rule-based controller.
  • ad hoc to paper Hand-crafted good/bad experience examples improve the LLM's decision quality
    The supplementary material provides author-chosen examples of high- and low-reward experiences. The representativeness of these examples across the state space is not analyzed, and their influence on performance is not ablated.
  • ad hoc to paper LLM output formatting is reliably parseable
    The method requires the LLM to wrap its decision in specific tags like <meta_action> and <tran_action>. No failure rate or fallback mechanism for malformed outputs is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical and Collaborative LLM-Based Control for Multi-UAV Motion and Communication in Integrated Terrestrial and Non-Terrestrial Networks." pith.science (2026). https://pith.science/paper/GFXT43R6

@misc{pith2026250606532,
  author       = {Pith},
  title        = {Pith review of: Hierarchical and Collaborative LLM-Based Control for Multi-UAV Motion and Communication in Integrated Terrestrial and Non-Terrestrial Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GFXT43R6}},
  note         = {Machine review of arXiv:2506.06532}
}
read the original abstract

Unmanned aerial vehicles (UAVs) have been widely adopted in various real-world applications. However, the control and optimization of multi-UAV systems remain a significant challenge, particularly in dynamic and constrained environments. This work explores the joint motion and communication control of multiple UAVs operating within integrated terrestrial and non-terrestrial networks that include high-altitude platform stations (HAPS). Specifically, we consider an aerial highway scenario in which UAVs must accelerate, decelerate, and change lanes to avoid collisions and maintain overall traffic flow. Different from existing studies, we propose a novel hierarchical and collaborative method based on large language models (LLMs). In our approach, an LLM deployed on the HAPS performs UAV access control, while another LLM onboard each UAV handles motion planning and control. This LLM-based framework leverages the rich knowledge embedded in pre-trained models to enable both high-level strategic planning and low-level tactical decisions. This knowledge-driven paradigm holds great potential for the development of next-generation 3D aerial highway systems. Experimental results demonstrate that our proposed collaborative LLM-based method achieves higher system rewards, lower operational costs, and significantly reduced UAV collision rates compared to baseline approaches.

Figures

Figures reproduced from arXiv: 2506.06532 by the authors.

Figure 1
Figure 1. Illustration of the proposed aerial network model. Assuming that one BS equipped with N antennas, and the array factor of the ULA of BS i and UAV j is F(ζ ij t ) = sin( Nπ 2 (sin ζ ij t − sin ζ d i )) √ N sin( π 2 (sin ζ ij t − sin ζ d i )) , (4) where ζ d i is the down-tilt of BS i’s ULA. Finally, the array radiation pattern from BS i to UAV j is B ij t = B(ζ ij t , ϕij t ) + F(ζ ij t ). (5) Then, the probability o… view at source ↗
Figure 2
Figure 2. Illustration of the proposed aerial network model (top view). Blue circles represent BSs; Solid/dash lines represent de￾sired/interference link. as (Alsharoa & Alouini, 2020): B Hj t = B  c 4π dHj t fc 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance Comparison 4. Numerical Results and Discussions 4.1. Experiment Setup We consider a scenario with a total of five UAVs (NUAV = 5) navigating a five-lane aerial highway, with flying speeds ranging from 5 m/s to 20 m/s. GBSs operate at a carrier frequency of 2.1 GHz and transmit at a power of PT = 40 dBm. Each BS supports a maximum of three UAV users and is deployed in varying quantities (5, 10, 15, or 20)… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Evaluation Comparison and peaks at a reward of 30. In contrast, DDQN saturates near 23, while Envelope-MHRL plateaus below 20. (b) Transportation Reward: A similar performance rank￾ing is observed. After convergence, the Dual-Agent model consistently achieves ∼25% high…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Intelligent Multi-UAV Navigation in ITNTNs: A Hierarchical LLM Approach

    cs.RO 2026-07 conditional novelty 4.0 of 10

    A cloud/edge LLM hierarchy that re-weights a fast UAV controller's reward function is claimed to reduce collisions and improve throughput, but the simulations do not directly plot those two headline metrics.

  2. Large Language Models for Next-Generation Wireless Network Management: A Survey and Tutorial

    cs.NI 2025-09 conditional novelty 4.0 of 10

    A survey and tutorial that organizes LLM-enabled wireless network optimization into formulation, solution, and verification stages, with case studies drawn from the authors' own prior papers.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages · cited by 2 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Study on enhanced LTE support for aerial vehicles (release 15), TR 36.777

    3GPP. Study on enhanced LTE support for aerial vehicles (release 15), TR 36.777. Jun. 2018

  3. [3]

    and Alouini, M.-S

    Alsharoa, A. and Alouini, M.-S. Improvement of the global connectivity using integrated satellite-airborne-terrestrial networks with resource optimization. IEEE Transactions on Wireless Communications, 19 0 (8): 0 5088--5100, 2020

  4. [4]

    Efficient drone mobility support using reinforcement learning

    Chen, Y., Lin, X., Khan, T., and Mozaffari, M. Efficient drone mobility support using reinforcement learning. In Proc. IEEE Wireless Commun. Network. Conf. (WCNC), pp.\ 1--6, 2020

  5. [5]

    Cellular-connected UAV in Next-Generation Wireless Networks

    Cherif, N. Cellular-connected UAV in Next-Generation Wireless Networks . PhD thesis, Universit \'e d'Ottawa/University of Ottawa, 2022

  6. [6]

    3d aerial highway: The key enabler of the retail industry transformation

    Cherif, N., Jaafar, W., Yanikomeroglu, H., and Yongacoglu, A. 3d aerial highway: The key enabler of the retail industry transformation. IEEE Communications Magazine, 59 0 (9): 0 65--71, 2021

  7. [7]

    Rl-based cargo-uav trajectory planning and cell association for minimum handoffs, disconnectivity, and energy consumption

    Cherif, N., Jaafar, W., Yanikomeroglu, H., and Yongacoglu, A. Rl-based cargo-uav trajectory planning and cell association for minimum handoffs, disconnectivity, and energy consumption. IEEE Transactions on Vehicular Technology, 2023

  8. [8]

    Design considerations for autonomous cargo transportation multirotor uavs

    Kotarski, D., Piljek, P., and Kasa \'c , J. Design considerations for autonomous cargo transportation multirotor uavs. Self-Driving Vehicles and Enabling Technologies, 2020

Show all 22 references
  1. [9]

    K., Khoshkholgh, M

    Kurt, G. K., Khoshkholgh, M. G., Alfattani, S., Ibrahim, A., Darwish, T. S., Alam, M. S., Yanikomeroglu, H., and Yongacoglu, A. A vision and framework for the high altitude platform station (haps) networks of the future. IEEE Communications Surveys & Tutorials, 23 0 (2): 0 729...

  2. [10]

    A novel uav-enabled data collection scheme for intelligent transportation system through uav speed control

    Li, X., Tan, J., Liu, A., Vijayakumar, P., Kumar, N., and Alazab, M. A novel uav-enabled data collection scheme for intelligent transportation system through uav speed control. IEEE Transactions on Intelligent Transportation Systems, 22 0 (4): 0 2100--2110, 2020

  3. [11]

    Large language model-based wireless network design

    Qiu, K., Bakirtzis, S., Wassell, I., Song, H., Zhang, J., and Wang, K. Large language model-based wireless network design. IEEE Wireless Communications Letters, 2024

  4. [12]

    K., Yanikomeroglu, H., and Chen, J

    Ren, Q., Abbasi, O., Kurt, G. K., Yanikomeroglu, H., and Chen, J. Handoff-aware distributed computing in high altitude platform station (haps)--assisted vehicular networks. IEEE Transactions on Wireless Communications, 2023

  5. [13]

    and Kesting, A

    Treiber, M. and Kesting, A. Traffic Flow Dynamics: Data Models and Simulation. Springer-Verlag Berlin Heidelberg, Sensors, 2013

  6. [14]

    I., and Letaief, K

    Xu, M., Niyato, D., Kang, J., Xiong, Z., Mao, S., Han, Z., Kim, D. I., and Letaief, K. B. When large language model agents meet 6g networks: Perception, grounding, and alignment. IEEE Wireless Communications, 2024

  7. [15]

    and Tabassum, H

    Yan, Z. and Tabassum, H. Reinforcement learning for joint v2i network selection and autonomous driving policies. In GLOBECOM 2022 - 2022 IEEE Global Communications Conference, pp.\ 1241--1246, 2022. doi:10.1109/GLOBECOM48099.2022.10001396

  8. [16]

    and Tabassum, H

    Yan, Z. and Tabassum, H. Generalized multi-objective reinforcement learning with envelope updates in urllc-enabled vehicular networks. arXiv preprint arXiv:2405.11331, 2024

  9. [17]

    Multi-uav speed control with collision avoidance and handover-aware cell association: Drl with action branching

    Yan, Z., Jaafar, W., Selim, B., and Tabassum, H. Multi-uav speed control with collision avoidance and handover-aware cell association: Drl with action branching. In GLOBECOM 2023-2023 IEEE Global Communications Conference, pp.\ 5067--5072. IEEE, 2023

  10. [18]

    Hybrid llm-ddqn-based joint optimization of v2i communication and autonomous driving

    Yan, Z., Zhou, H., Tabassum, H., and Liu, X. Hybrid llm-ddqn-based joint optimization of v2i communication and autonomous driving. IEEE Wireless Communications Letters, 14 0 (4): 0 1214--1218, 2025. doi:10.1109/LWC.2025.3539638

  11. [19]

    Deep learning for channel tracking in irs-assisted uav communication systems

    Yu, J., Liu, X., Gao, Y., Zhang, C., and Zhang, W. Deep learning for channel tracking in irs-assisted uav communication systems. IEEE Transactions on Wireless Communications, 21 0 (9): 0 7711--7722, 2022

  12. [20]

    Game of drones: Multi-uav pursuit-evasion game with online motion planning by deep reinforcement learning

    Zhang, R., Zong, Q., Zhang, X., Dou, L., and Tian, B. Game of drones: Multi-uav pursuit-evasion game with online motion planning by deep reinforcement learning. IEEE Transactions on Neural Networks and Learning Systems, 2022

  13. [21]

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

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

  14. [22]

    Large language models for wireless networks: An overview from the prompt engineering perspective

    Zhou, H., Hu, C., Yuan, D., Yuan, Y., Wu, D., Chen, X., Tabassum, H., and Liu, X. Large language models for wireless networks: An overview from the prompt engineering perspective. IEEE Wireless Communications, 2025

Pith tools

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