Pith. sign in

REVIEW 5 major objections 5 minor 57 references

Onboard vision-language inference on a UAV can be made efficient by choosing per-user image resolution and transmit power at session level and letting an LLM-refined, risk-aware reward guide the drone's trajectory, cutting worst-case task l

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 10:21 UTC pith:EPLA5QHR

load-bearing objection A plausible framework for resolution–power–trajectory optimization in UAV VLM inference, but the central KKT-based optimality proof has a load-bearing gap that undercuts the numerical claims until fixed. the 5 major comments →

arxiv 2510.10028 v2 pith:EPLA5QHR submitted 2025-10-11 cs.LG cs.AIcs.DC

Efficient Onboard Vision-Language Inference in UAV-Enabled Low-Altitude Economy Networks via LLM-Enhanced Optimization

Yang Li , Ruichen Zhang , Yinqiu Liu , Guangyuan Liu , Abbas Jamalipour , Xianbin Wang , Dong In Kim This is my paper
classification cs.LG cs.AIcs.DC
keywords low-altitude economy networksUAV trajectory optimizationvision-language model inferenceresource allocationLLM reward designdeep reinforcement learningmixed-integer programmingtask latency minimization
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 UAV serving vision-language inference to ground users can jointly manage three coupled costs — uplink image payload, onboard VLM processing, and flight path — without sacrificing per-user accuracy. Its central claim is that the mixed-integer problem of choosing discrete image resolutions, continuous transmit powers, and a 3D trajectory can be split into a solver-based session-level allocation (ARPO) and a reinforcement-learning trajectory controller (LLaRA) whose reward is designed offline by a large language model. Under a constant-rate approximation for each upload session, ARPO provably gives the optimal resolution–power pair; LLaRA's LLM-refined tail-aware reward then steers the UAV toward bottleneck users and reduces worst-case latency by about 11.7–13.7% over a manually rewarded PPO baseline. A sympathetic reader would care because this offers a concrete route to “inference-as-a-service” from aerial platforms, where communication and AI accuracy are usually optimized separately.

Core claim

On the paper's own terms, the discovery is that the accuracy–efficiency trade-off of vision-language models is an exploitable degree of freedom for aerial networks. Image resolution saturates in accuracy, so the smallest resolution that meets each user's accuracy constraint dominates higher ones in both payload and processing time; this makes resolution selection a lightweight branch-and-bound lookup. Given that resolution, power allocation under a fixed-session channel becomes convex, with a closed-form KKT solution and a one-dimensional bisection. The remaining trajectory problem is solved by PPO, but instead of a hand-crafted penalty the agent maximizes an LLM-generated reward that is the

What carries the argument

Two coupled mechanisms carry the argument. (1) ARPO: a lookup table of measured accuracy and inference speed per resolution converts the accuracy constraint into a finite search; branch-and-bound picks the minimal feasible resolution, and after substituting it the power subproblem is convex, so KKT conditions give power as a closed-form function of a single latency parameter, found by bisection. (2) LLaRA: a large language model, prompted with the system model and evaluation feedback, generates and iteratively refines candidate reward functions; the selected “risk-aware” reward is the negative empirical quantile of unfinished data plus a throughput term and a distance-to-bottleneck-user term

Load-bearing premise

The load-bearing premise is that each user's image resolution and transmit power stay fixed for the whole upload session and that the uplink channel can be evaluated once at the session's start as a constant rate; if the channel changes materially during an upload, the claimed optimality of the resolution–power allocation no longer holds.

What would settle it

Run the same system with strong per-slot small-scale fading while the drone is moving, and compare ARPO's fixed session-level resolution and power choice against a version that re-optimizes power each slot: if per-slot re-optimization cuts worst-case latency by more than a small margin, the constant-rate assumption behind ARPO's optimality is violated.

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

If this is right

  • The smallest resolution meeting an accuracy requirement is always preferable for a fixed power budget, since accuracy gains saturate while payload and inference time grow.
  • A mixed-integer, non-convex joint allocation problem can be decomposed into a convex power subproblem and a DRL trajectory subproblem, turning an NP-hard problem into tractable components.
  • Using a tail-risk (quantile) reward instead of a mean-backlog reward makes the learned trajectory align with worst-case latency rather than average performance.
  • LLM-refined reward design improves both convergence speed and final latency relative to manually designed rewards, with zero added latency during deployment.
  • The same allocation choices generalize across user batches, allowing the UAV to continue serving successive groups without returning to base.

Where Pith is reading between the lines

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

  • Because ARPO relies on a constant-rate approximation, the framework's edge is largest when channels change slowly; a slot-level re-optimization or predictive channel model would be a natural extension the authors do not explore.
  • The offline LLM reward loop is transferable to other multi-user scheduling MDPs: any system with a straggler or tail objective could adopt the quantile-plus-throughput-plus-distance reward template.
  • A testable extension would be to rerun the reward loop with a different vision-language model or benchmark; the expected gain depends on how strongly that model's accuracy saturates with resolution, a claim demonstrated only for the one reference model in the paper.
  • The multi-round batch experiment hints at a lifelong service formulation, but the paper does not formalize learning across batches; the LLM reward could be re-refined as user distributions shift.

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

5 major / 5 minor

Summary. The paper proposes a UAV-enabled Low-Altitude Economy Network in which a single UAV runs an onboard vision-language model (VLM) to serve multiple ground users. It formulates P0 (Eq. 21), a mixed-integer non-convex problem minimizing worst-case task latency plus weighted transmit power under per-user accuracy constraints, discrete resolution choices, mobility constraints, and power caps. The solution is hierarchical: ARPO (Branch-and-Bound over resolution, KKT-based closed form for power, Proposition 1, Eqs. 27-31, Algorithm 1) and LLaRA (PPO with an offline LLM-refined reward function, Eqs. 37-40). Accuracy and inference-speed lookup tables (Eqs. 45-46) come from empirical profiling of LLaVA-HR on TextVQA. Numerical results report roughly 11.7%-13.7% latency improvement over an ARPO-PPO baseline with a manually designed reward (Section VII-C, Figs. 7-8).

Significance. If the technical gaps are fixed, the paper would contribute a useful system-level model coupling the VLM resolution-accuracy-efficiency trade-off with A2G communication and UAV trajectory planning, and a concrete application of offline LLM reward design that introduces no online inference latency. Strengths worth noting: the LLaVA-HR profiling data (Eqs. 45-46) are external empirical inputs rather than fitted outputs of the optimizer; the hierarchical decomposition is plausible; the ablations over zeta, bandwidth, Pmax, and the multi-round scenario are informative. However, the central ARPO optimality proof is incomplete as written, and the reported percentage improvements are internally inconsistent, so the headline claims are not yet established.

major comments (5)
  1. [§IV, Proposition 1, Eqs. (27)-(31)] The closed form omits the power-cap regime. When P_n=P_max is active, iota_n>0 and stationarity (27a) gives omega_n=(zeta+iota_n)/g_n > zeta/g_n; the proof's step 'with iota_n=0' is unjustified, so Eq. (30) is not the correct stationarity condition in that regime. Counterexample (D1=10, D2=1, h1=h2=1, sigma^2=1, B=1, P_max=1, zeta=1): the true optimum is tau*=10, P1=1 (cap active), P2=2^0.1-1≈0.072; for all feasible tau>=10 the LHS of (30) is <1 (≈0.146 at tau=10), so (30) has no feasible root, and its root at tau<10 gives clamped powers violating (25a). Algorithm 1 is therefore not a correct solver of P2 as stated; an active-set procedure or a verified no-binding condition plus feasibility check is required. Also, the text calls g_n in (28) 'the derivative' of T_total, but it is the magnitude; the derivative is -g_n.
  2. [§IV (text before Algorithm 1) vs. Eq. (46)] The paper asserts 'the inference speed function v(r_n) is non-decreasing with resolution', which is directly contradicted by its own profiling table (46): v drops from 23.8 to 12.6 tokens/s as r_n grows from 384p to 1536p. The conclusion that the smallest feasible resolution is optimal does survive, but only because D(r_n) in (7) and T_proc=E[|A_pred|]/v(r_n) in (17) are both non-decreasing in r_n; the stated premise is false and the monotonicity argument must be rewritten, e.g., using that 1/v(r_n) and D(r_n) are non-decreasing.
  3. [§III-C, Assumption 1, and Eq. (22)] ARPO's optimality is derived under a channel frozen at the starting pose, replacing slot-varying rates with the constant R_n. The simulation parameters undermine the assumption's justification: with alpha=1 s, V_xy,max=100 m/s, and upload windows of ~10-15 s (Figs. 7-8), the UAV travels up to ~1.5 km while user distances are ~1 km, so the A2G gain (beta0/d^gamma, gamma=2) changes by orders of magnitude during a single upload. The simulator evaluates latency via the slot-varying model (Eqs. 5-11), so the experiments do not validate ARPO optimality for the actual P0. Please quantify the approximation error with the given parameters, or state clearly that optimality holds only for the quasi-static surrogate.
  4. [§V-B and §VII-C] The LLM-designed rewards are scored and refined by evaluating candidates in the same latency objective the reward is meant to optimize (Eqs. 37-40), whereas the ARPO-PPO baseline uses a single manually crafted reward with no selection or tuning budget. The claimed 11.7%-13.7% gain may thus reflect the search/refinement budget rather than the LLM's design ability. A control is needed: e.g., apply the same iterative selection to K randomly generated or hand-written rewards, or grant the manual reward an equal number of refinement rounds, before attributing the gain to the LLM.
  5. [§VII-C and Figs. 7-8] The reported percentages are internally inconsistent. Fig. 7(b) labels (29.7%, 10.7%, 12.2%) match the plotted values (22.65->15.92->14.21->12.48), but the surrounding text for Fig. 7 reports 27.4%, 10.8%, 11.7% - the same numbers that correctly describe Fig. 8(b) (19.73->14.32->12.78->11.29). The 13.7% improvement claimed in the convergence discussion appears in no table or figure. The headline range '11.7%-13.7%' must be recomputed consistently from a single source of truth.
minor comments (5)
  1. [§III-A] Typo: 'Our system model be readily extended' should read 'can be readily extended'.
  2. [Eq. (3a) vs. Eq. (41)] d_n[t] denotes the UAV-user distance in Eq. (3a), but is reused for remaining data size in Eq. (41) and Eq. (44). Use a distinct symbol (e.g., d_rem,n[t]) to avoid ambiguity.
  3. [§VII-A] The default value of zeta used in Figs. 6-8 and 10-11 is not stated; zeta enters the ARPO solution directly via Eq. (30) and is needed for reproducibility.
  4. [Eq. (28)] The text says g_n is the derivative of T_total w.r.t. P_n, but the expression is positive, i.e., it is the absolute value of that derivative (the derivative itself is negative). The KKT condition (27a) is consistent with the magnitude convention, so only the wording needs correction.
  5. [§V-B] The chain-of-thought prompting strategy is cited as [1]; the canonical reference is Wei et al., 2022. Also verify the model name 'Qwen-3-225B' in Fig. 5 against reference [56].

Circularity Check

0 steps flagged

No significant circularity: external lookup tables are inputs, ARPO/LLaRA solve well-defined problems; minor self-citations are not load-bearing.

full rationale

The claimed derivation chain is not circular by construction. Accuracy and speed mappings A_n(r_n) and v(r_n) in Eqs. (45)-(46) are externally profiled LLaVA-HR values used as inputs to P0; the optimizer does not produce them. The reductions P0->P1 (epigraph variable tau) and P1->P2 (fix r, convexify in P) are standard transformations, and Proposition 1 is a KKT solution of the stated convex subproblem. The reviewer's concrete counterexample targets an omitted active power-bound case in Eq. (30); that is a mathematical correctness/omitted-proof issue, not a definitional circularity. LLaRA's LLM reward loop scores candidate rewards in the DRL environment and selects the best (Eq. 40); reporting the trained policy's latency is reporting the result of that search. If no held-out environment is used, this raises model-selection-on-the-evaluation-metric concerns, but it is not a fitted parameter being renamed as a prediction, because the final latency is not an independent claim derived from the scores. Self-citations [1] and [14] are used for background and for standard RL/LLM context; the central objective, ARPO solution, and trajectory MDP do not rest on them, and no author-specific uniqueness theorem is invoked. The paper's statement that v(r_n) is non-decreasing is inconsistent with its own lookup table (46), but this is also a correctness/typo issue, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim rests on empirical lookup tables fitted to a single VLM, a quasi-static channel assumption that lets ARPO use closed-form rates, and the empirical hypothesis that LLM-selected rewards outperform hand-designed rewards. No new physical or conceptual entities are invented; the method reuses existing components.

free parameters (6)
  • A_n(r_n) accuracy lookup table = 59.63%, 64.36%, 67.11%, 67.96% at 384/768/1024/1536p
    Empirical TextVQA accuracy of LLaVA-HR (Eq. 45) used as the exact accuracy function in constraint C1; no error bars or per-user variation.
  • v(r_n) inference speed lookup table = 23.8, 19.9, 19.7, 12.6 tokens/s at 384/768/1024/1536p
    Profiled inference speed (Eq. 46) used to compute VLM processing time; decreases with resolution, contradicting the monotonicity claim in §IV.
  • ζ latency-power weight = 100 to 1000 in experiments
    Tunable coefficient in objective (21); no principled selection rule is given.
  • LLM reward weights (q, μ, γ_d) = not specified
    Weights in the LLM reward Eq. (44) are never assigned in the experiments; policy behavior depends on them.
  • Expected answer length E[|A_pred|] = not stated
    Required by Eq. (17) to compute T_proc; no value or distribution is given, so latency numbers implicitly depend on an unstated quantity.
  • Channel/system parameter set = a=4.88, b=0.43, γ_LoS=γ_NLoS=2, β0=-50 dB, σ²=-100 dBm, B_n=1 MHz, P_max=0.1 W
    Adapted from prior literature and hand-chosen for simulation; all latency and power results depend on these values.
axioms (6)
  • standard math Problem P2 is convex and KKT conditions give its global optimum
    Used in Section IV to derive Proposition 1; relies on the convexity claim for the power-only subproblem.
  • domain assumption Elevation-dependent LoS/NLoS probabilistic air-to-ground channel model
    Eq. (4) and (5) define the channel; the whole communication model is built on this standard assumption.
  • domain assumption Quasi-static channels and session-fixed (r_n, P_n) with constant rate during upload
    Assumption 1 and Eq. (22) replace slot-varying rates with a constant R_n; this is load-bearing for ARPO's closed-form latency and optimality.
  • domain assumption The LLaVA-HR lookup tables for A_n(r_n) and v(r_n) generalize to all users and deployment conditions
    Eqs. (45)–(46) are used as exact inputs to C1 and T_proc for every user; no validation on other models, hardware, or environments.
  • ad hoc to paper v(r_n) is non-decreasing with resolution (§IV)
    This stated premise is contradicted by the paper's own Eq. (46), where speed drops from 23.8 to 12.6 tokens/s as resolution increases. The monotonicity conclusion may still hold for other reasons, but the text's justification is invalid.
  • ad hoc to paper LLM-generated reward functions evaluated in the simulator yield better and more stable policies than manually designed rewards
    The core hypothesis of LLaRA is supported only by unreleased simulations; no LLM output transcripts, candidate list, or independent validation are provided.

pith-pipeline@v1.3.0-alltime-deepseek · 23108 in / 15739 out tokens · 146776 ms · 2026-08-04T10:21:08.752270+00:00 · methodology

0 comments
read the original abstract

The rapid advancement of Low-Altitude Economy Networks (LAENets) has enabled a variety of applications, including aerial surveillance, environmental sensing, and semantic data collection. To support these scenarios, unmanned aerial vehicles (UAVs) equipped with onboard vision-language models (VLMs) offer a promising solution for real-time multimodal inference. However, ensuring both inference accuracy and communication efficiency remains a significant challenge due to limited onboard resources and dynamic network conditions. In this paper, we first propose a UAV-enabled LAENet system model that jointly captures UAV mobility, user-UAV communication, and the onboard visual question answering (VQA) pipeline. Based on this model, we formulate a mixed-integer non-convex optimization problem to minimize task latency and power consumption under user-specific accuracy constraints. To solve the problem, we design a hierarchical optimization framework composed of two parts: (i) an Alternating Resolution and Power Optimization (ARPO) algorithm for resource allocation under accuracy constraints, and (ii) a Large Language Model-augmented Reinforcement Learning Approach (LLaRA) for adaptive UAV trajectory optimization. The large language model (LLM) serves as an expert in refining reward design of reinforcement learning in an offline fashion, introducing no additional latency in real-time decision-making. Numerical results demonstrate the efficacy of our proposed framework in improving inference performance and communication efficiency under dynamic LAENet conditions.

Figures

Figures reproduced from arXiv: 2510.10028 by Abbas Jamalipour, Dong In Kim, Guangyuan Liu, Ruichen Zhang, Xianbin Wang, Yang Li, Yinqiu Liu.

Figure 1
Figure 1. Figure 1: An overview of the onboard VLM inference-driven [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The framework of our proposed hierarchical ARPO-LLaRA optimization framework. At the start of uplink session, ARPO [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The workflow of LLaRA. The LLM-augmented reward design employs an LLM expert to generate and iteratively refine [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Instance prompts used in the initialization and evolution [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of input resolution on TextVQA. Left and middle: our reproduced experiments using LLaVA-HR show that [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Convergence performance comparison between ARPO [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overall latency and sum of user transmit powers under [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Optimized UAV trajectory and minimum latency com [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: UAV trajectory performance for a multi-round service scenario. The UAV starts from the initial location, sequentially [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The overall latency versus different maximum transmit [PITH_FULL_IMAGE:figures/full_fig_p015_11.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

57 extracted references · 22 linked inside Pith

  1. [1]

    Large language model-enhanced reinforcement learning for low-altitude economy networking,

    L. Cai, R. Zhang, C. Zhao, Y . Zhang, J. Kang, D. Niyato, T. Jiang, and X. Shen, “Large language model-enhanced reinforcement learning for low-altitude economy networking,”arXiv preprint arXiv:2505.21045, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Satellite-assisted low-altitude economy networking: Concepts, applications, and opportu- nities,

    S. He, J. Wang, Y .-C. Liang, G. Sun, and D. Niyato, “Satellite-assisted low-altitude economy networking: Concepts, applications, and opportu- nities,”arXiv preprint arXiv:2505.04098, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Secure physical layer communica- tions for low-altitude economy networking: A survey,

    L. Cai, J. Wang, R. Zhang, Y . Zhang, T. Jiang, D. Niyato, X. Wang, A. Jamalipour, and X. Shen, “Secure physical layer communica- tions for low-altitude economy networking: A survey,”arXiv preprint arXiv:2504.09153, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Toward realization of low-altitude economy networks: Core architecture, integrated technologies, and future directions,

    Y . Wang, G. Sun, Z. Sun, J. Wang, J. Li, C. Zhao, J. Wu, S. Liang, M. Yin, P. Wanget al., “Toward realization of low-altitude economy networks: Core architecture, integrated technologies, and future directions,”arXiv preprint arXiv:2504.21583, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    China’s low-altitude economy soars at high speed,

    C. Daily, “China’s low-altitude economy soars at high speed,” https: //www.chinadaily.com.cn/a/202412/19/WS6763b8b7a310f1265a1d3d24. html, 2024

    2024

  6. [6]

    Ubiquitous uav communication enabled low-altitude economy: Applications, techniques, and 3gpp’s efforts,

    D. He, W. Yuan, J. Wu, and R. Liu, “Ubiquitous uav communication enabled low-altitude economy: Applications, techniques, and 3gpp’s efforts,”IEEE Network, 2025

    2025

  7. [7]

    Emerging uav technology for disaster detection, mitigation, response, and preparedness,

    A. Khan, S. Gupta, and S. K. Gupta, “Emerging uav technology for disaster detection, mitigation, response, and preparedness,”Journal of Field Robotics, vol. 39, no. 6, pp. 905–955, 2022

    2022

  8. [8]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,”Advances in neural information processing systems, vol. 36, 2023

    2023

  9. [9]

    A vision– language model-based traffic sign detection method for high-resolution drone images: A case study in guyuan, china,

    J. Yao, J. Li, Y . Li, M. Zhang, C. Zuo, S. Dong, and Z. Dai, “A vision– language model-based traffic sign detection method for high-resolution drone images: A case study in guyuan, china,”Sensors, 2024

    2024

  10. [10]

    Clip-ug: Clip-driven vision-language model for uav-view geo-localization,

    J. Wu and G. Feng, “Clip-ug: Clip-driven vision-language model for uav-view geo-localization,”IEEE Transactions on Consumer Electronics, 2025

    2025

  11. [11]

    Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” inInternational conference on machine learning. PMLR, 2023

    2023

  12. [12]

    General-purpose aerial intelligent agents empowered by large language models,

    J. Zhao and X. Lin, “General-purpose aerial intelligent agents empowered by large language models,”arXiv preprint arXiv:2503.08302, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025. 16

    Pith/arXiv arXiv 2025

  14. [14]

    Embodied ai-enhanced vehicular networks: An integrated vision language models and reinforcement learning method,

    R. Zhang, C. Zhao, H. Du, D. Niyato, J. Wang, S. Sawadsitang, X. Shen, and D. I. Kim, “Embodied ai-enhanced vehicular networks: An integrated vision language models and reinforcement learning method,” IEEE Transactions on Mobile Computing, pp. 1–16, 2025

    2025

  15. [15]

    Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models,

    G. Luo, Y . Zhou, Y . Zhang, X. Zheng, X. Sun, and R. Ji, “Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models,” arXiv preprint arXiv:2403.03003, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    6g non-terrestrial networks enabled low-altitude economy: Opportunities and challenges,

    Y . Jiang, X. Li, G. Zhu, H. Li, J. Deng, K. Han, C. Shen, Q. Shi, and R. Zhang, “6g non-terrestrial networks enabled low-altitude economy: Opportunities and challenges,”arXiv preprint arXiv:2311.09047, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Toward a sustainable low-altitude economy: A survey of energy- efficient ris-uav networks,

    M. Ahmed, A. A. Soofi, F. Khan, S. Raza, W. U. Khan, L. Su, F. Xu, and Z. Han, “Toward a sustainable low-altitude economy: A survey of energy- efficient ris-uav networks,”arXiv preprint arXiv:2504.02162, 2025

    arXiv 2025

  18. [18]

    Energy-efficient irreg- ular ris-aided uav-assisted optimization: A deep reinforcement learning approach,

    M. M. Salim, K. M. Rabie, and A. H. Muqaibel, “Energy-efficient irreg- ular ris-aided uav-assisted optimization: A deep reinforcement learning approach,”arXiv preprint arXiv:2504.15031, 2025

    arXiv 2025

  19. [19]

    Embodied ai- empowered low altitude economy: Integrated sensing, communications, computation, and control (isc3),

    Y . Yang, Y . Chen, J. Wang, G. Sun, and D. Niyato, “Embodied ai- empowered low altitude economy: Integrated sensing, communications, computation, and control (isc3),”arXiv preprint arXiv:2412.19996, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Unmanned aerial vehicles based low-altitude economy with lifecycle techno-economic-environmental analysis for sustainable and smart cities,

    Y . Zhou, “Unmanned aerial vehicles based low-altitude economy with lifecycle techno-economic-environmental analysis for sustainable and smart cities,”Journal of Cleaner Production, p. 145050, 2025

    2025

  21. [21]

    Generative AI agents with large language model for satellite networks via a mixture of experts transmission,

    R. Zhang, H. Du, Y . Liu, D. Niyato, J. Kang, Z. Xiong, A. Jamalipour, and D. In Kim, “Generative AI agents with large language model for satellite networks via a mixture of experts transmission,”IEEE Journal on Selected Areas in Communications, vol. 42, no. 12, 2024

    2024

  22. [22]

    Patrol agent: An autonomous uav framework for urban patrol using on board vision language model and on cloud large language model,

    Z. Yuan, F. Xie, and T. Ji, “Patrol agent: An autonomous uav framework for urban patrol using on board vision language model and on cloud large language model,” in2024 6th International Conference on Robotics and Computer Vision (ICRCV). IEEE, 2024, pp. 237–242

    2024

  23. [23]

    Multimodal ai for uav: Vision– language models in human–machine collaboration,

    M. Krup ´aˇs, L. Urbl ´ık, and I. Zolotov ´a, “Multimodal ai for uav: Vision– language models in human–machine collaboration,”Electronics, 2025

    2025

  24. [24]

    Vision-language models for edge networks: A comprehensive survey,

    A. Sharshar, L. U. Khan, W. Ullah, and M. Guizani, “Vision-language models for edge networks: A comprehensive survey,”IEEE Internet of Things Journal, 2025

    2025

  25. [25]

    Litevlm: A low-latency vision- language model inference pipeline for resource-constrained environ- ments,

    J. Huang, Y . Jin, L. An, and J. Park, “Litevlm: A low-latency vision- language model inference pipeline for resource-constrained environ- ments,”arXiv preprint arXiv:2506.07416, 2025

    arXiv 2025

  26. [26]

    Disco: Device-server collaborative llm- based text streaming services,

    T. Sun, P. Wang, and F. Lai, “Disco: Device-server collaborative llm- based text streaming services,”arXiv preprint arXiv:2502.11417, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Perllm: Person- alized inference scheduling with edge-cloud collaboration for diverse llm services,

    Z. Yang, Y . Yang, C. Zhao, Q. Guo, W. He, and W. Ji, “Perllm: Person- alized inference scheduling with edge-cloud collaboration for diverse llm services,”arXiv preprint arXiv:2405.14636, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Distributed vlms: Efficient vision-language processing through cloud- edge collaboration,

    Y . Li, D. Gumaste, M. K. Turkcan, J. Ghaderi, G. Zussman, and Z. Kostic, “Distributed vlms: Efficient vision-language processing through cloud- edge collaboration,” in2025 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops). IEEE Computer Society, 2025, pp. 280–286

    2025

  29. [29]

    Energy efficiency maximization in ris-assisted swipt networks with rsma: A ppo-based approach,

    R. Zhang, K. Xiong, Y . Lu, P. Fan, D. W. K. Ng, and K. B. Letaief, “Energy efficiency maximization in ris-assisted swipt networks with rsma: A ppo-based approach,”IEEE Journal on Selected Areas in Communi- cations, vol. 41, no. 5, pp. 1413–1430, 2023

    2023

  30. [30]

    Survey on large language model-enhanced reinforce- ment learning: Concept, taxonomy, and methods,

    Y . Cao, H. Zhao, Y . Cheng, T. Shu, Y . Chen, G. Liu, G. Liang, J. Zhao, J. Yan, and Y . Li, “Survey on large language model-enhanced reinforce- ment learning: Concept, taxonomy, and methods,”IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  31. [31]

    Natural language- conditioned reinforcement learning with inside-out task language devel- opment and translation,

    J.-C. Pang, X.-Y . Yang, S.-H. Yang, and Y . Yu, “Natural language- conditioned reinforcement learning with inside-out task language devel- opment and translation,”arXiv preprint arXiv:2302.09368, 2023

    Pith/arXiv arXiv 2023

  32. [32]

    Informing reinforcement learning agents by grounding language to markov decision processes,

    B. A. Spiegel, Z. Yang, W. Jurayj, B. Bachmann, S. Tellex, and G. Konidaris, “Informing reinforcement learning agents by grounding language to markov decision processes,” 2024

    2024

  33. [33]

    Reward design with language models,

    M. Kwon, S. M. Xie, K. Bullard, and D. Sadigh, “Reward design with language models,”arXiv preprint arXiv:2303.00001, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    Eureka: Human-level reward design via coding large language models,

    Y . J. Ma, W. Liang, G. Wang, D.-A. Huang, O. Bastani, D. Jayaraman, Y . Zhu, L. Fan, and A. Anandkumar, “Eureka: Human-level reward design via coding large language models,”arXiv preprint arXiv:2310.12931, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Text2reward: Reward shaping with language models for reinforcement learning,

    T. Xie, S. Zhao, C. H. Wu, Y . Liu, Q. Luo, V . Zhong, Y . Yang, and T. Yu, “Text2reward: Reward shaping with language models for reinforcement learning,”arXiv preprint arXiv:2309.11489, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Pre-trained language models for interactive decision-making,

    S. Li, X. Puig, C. Paxton, Y . Du, C. Wang, L. Fan, T. Chen, D.-A. Huang, E. Aky ¨urek, A. Anandkumaret al., “Pre-trained language models for interactive decision-making,”Advances in Neural Information Processing Systems, vol. 35, pp. 31 199–31 212, 2022

    2022

  37. [37]

    Option discovery using llm-guided semantic hierarchical reinforcement learning,

    C. L. Shek and P. Tokekar, “Option discovery using llm-guided semantic hierarchical reinforcement learning,”arXiv preprint arXiv:2503.19007, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Transformer- based world models are happy with 100k interactions,

    J. Robine, M. H ¨oftmann, T. Uelwer, and S. Harmeling, “Transformer- based world models are happy with 100k interactions,”arXiv preprint arXiv:2303.07109, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Towards balancing preference and performance through adaptive personalized explainability,

    A. Silva, P. Tambwekar, M. Schrum, and M. Gombolay, “Towards balancing preference and performance through adaptive personalized explainability,” inProceedings of the 2024 ACM/IEEE international conference on human-robot interaction, 2024, pp. 658–668

    2024

  40. [40]

    Completion time minimization for multi-uav- enabled data collection,

    C. Zhan and Y . Zeng, “Completion time minimization for multi-uav- enabled data collection,”IEEE Transactions on Wireless Communications, vol. 18, no. 10, pp. 4859–4872, 2019

    2019

  41. [41]

    Towards vqa models that can read,

    A. Singh, V . Natarajan, M. Shah, Y . Jiang, X. Chen, D. Batra, D. Parikh, and M. Rohrbach, “Towards vqa models that can read,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 8317–8326

    2019

  42. [42]

    Joint resource allocation and 3d aerial trajectory design for video streaming in uav communication systems,

    C. Zhan, H. Hu, X. Sui, Z. Liu, J. Wang, and H. Wang, “Joint resource allocation and 3d aerial trajectory design for video streaming in uav communication systems,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 8, pp. 3227–3241, 2020

    2020

  43. [43]

    Near optimum power control and precoding under fairness constraints in network mimo systems,

    G. Fodor, M. Johansson, and P. Soldati, “Near optimum power control and precoding under fairness constraints in network mimo systems,” International Journal of Digital Multimedia Broadcasting, vol. 2010, no. 1, p. 251719, 2010

    2010

  44. [44]

    Optimum power control at finite blocklength,

    W. Yang, G. Caire, G. Durisi, and Y . Polyanskiy, “Optimum power control at finite blocklength,”IEEE Transactions on Information Theory, vol. 61, no. 9, pp. 4598–4615, 2015

    2015

  45. [45]

    Energy-efficient multi-uavs cooperative trajectory optimization for communication cover- age: An madrl approach,

    T. Ao, K. Zhang, H. Shi, Z. Jin, Y . Zhou, and F. Liu, “Energy-efficient multi-uavs cooperative trajectory optimization for communication cover- age: An madrl approach,”Remote Sensing, vol. 15, no. 2, p. 429, 2023

    2023

  46. [46]

    Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images,

    Z. Guo, R. Xu, Y . Yao, J. Cui, Z. Ni, C. Ge, T.-S. Chua, Z. Liu, and G. Huang, “Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 390–406

    2024

  47. [47]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  48. [48]

    Resource allocation for the training of image semantic communication networks,

    Y . Li, X. Zhou, and J. Zhao, “Resource allocation for the training of image semantic communication networks,”IEEE Transactions on Wireless Communications, 2025

    2025

  49. [49]

    A branch and bound algorithm for transmission network expansion plan- ning using nonconvex mixed-integer nonlinear programming models,

    R. T. Zoppei, M. A. Delgado, L. H. Macedo, M. J. Rider, and R. Romero, “A branch and bound algorithm for transmission network expansion plan- ning using nonconvex mixed-integer nonlinear programming models,” IEEE Access, vol. 10, pp. 39 875–39 888, 2022

    2022

  50. [50]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  51. [51]

    Survey of hallucination in natural language generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . J. Bang, A. Madotto, and P. Fung, “Survey of hallucination in natural language generation,”ACM computing surveys, vol. 55, no. 12, pp. 1–38, 2023

    2023

  52. [52]

    Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning,

    D. R. Morrison, S. H. Jacobson, J. J. Sauppe, and E. C. Sewell, “Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning,”Discrete Optimization, vol. 19, 2016

    2016

  53. [53]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radfordet al., “Gpt-4o system card,”arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  54. [54]

    Erfsl: An efficient reward function searcher via large language models for custom-environment multi-objective reinforcement learning (student abstract),

    G. Xie, J. Xu, Y . Yang, Y . Ding, and S. Zhang, “Erfsl: An efficient reward function searcher via large language models for custom-environment multi-objective reinforcement learning (student abstract),” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 28, 2025

    2025

  55. [55]

    Monitoring reasoning models for misbehavior and the risks of promoting obfuscation,

    B. Baker, J. Huizinga, L. Gao, Z. Dou, M. Y . Guan, A. Madry, W. Zaremba, J. Pachocki, and D. Farhi, “Monitoring reasoning models for misbehavior and the risks of promoting obfuscation,”arXiv preprint arXiv:2503.11926, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  57. [57]

    Resilient mobile multi-target surveillance using multi-hop autonomous uav networks for extended lifetime,

    A. Da ˘gas ¸an and E. Karas ¸an, “Resilient mobile multi-target surveillance using multi-hop autonomous uav networks for extended lifetime,”arXiv preprint arXiv:2311.03030, 2023

    Pith/arXiv arXiv 2023