Pith. sign in

REVIEW 4 major objections 9 minor 25 references

LLM-guided UAVs reach 91% of optimal coverage in 6G AI service delivery

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 · glm-5.2

2026-07-08 11:00 UTC pith:LSGZSMVN

load-bearing objection HyPE combines DRL mobility prediction, LLM-guided UAV trajectory/inference assignment, and greedy heuristics for QoAIS-aware service provisioning in UAV-assisted 6G. The framework is architecturally novel but headline metrics in the conclusion are not clearly derivable from reported results. the 4 major comments →

arxiv 2607.06278 v1 pith:LSGZSMVN submitted 2026-07-07 cs.NI

Quality-Aware Personalized AI Service Provisioning in UAV-Assisted 6G Networks

Mohammad Farhoudi , Masoud Shokrnezhad , Tarik Taleb This is my paper
classification cs.NI
keywords servicequalityhypepersonalizedprovisioningconventionallatencylearning-augmented
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 introduces HyPE, a three-phase framework for orchestrating personalized AI services across UAV-assisted 6G networks. The central problem is that jointly optimizing UAV trajectories, inference node assignments, and service routing while balancing latency, output fidelity, and personalization continuity is NP-hard and computationally intractable at scale. HyPE decomposes the problem into three stages: a deep reinforcement learning predictor (MAP) forecasts where users will be and what they will request; a prompted large language model (LEAD) uses structured reasoning with role prompting, contrastive few-shot examples, and enforced output schemas to decide UAV movements and inference assignments; and deterministic heuristics (SET) handle final function placement and routing. The paper formulates the full optimization problem with 18 constraints covering trajectory adjacency, network topology, latency, fidelity, personalization continuity, and resource capacity, then demonstrates through simulation that HyPE achieves 91% of the oracle optimizer's coverage and 94% of optimal fidelity while maintaining interactive-grade latency under increasing user density.

Core claim

An LLM prompted with structured domain rules, contrastive few-shot examples, and machine-readable output schemas can serve as a real-time decision engine for combinatorial network orchestration problems, producing feasible UAV trajectory and inference assignment decisions that approach optimal quality without exhaustive enumeration. The paper shows this LLM-guided approach, combined with DRL-based demand prediction and deterministic heuristics for placement and routing, achieves 91% of optimal coverage and 94% of optimal fidelity in simulations using real mobility traces and heterogeneous AI workloads.

What carries the argument

The HyPE pipeline: (1) MAP uses Dueling Double Deep Q-Learning with LSTM-CNN at each point of attachment to predict user location and service intent distributions; (2) LEAD queries an LLM (Gemini-2.5-Flash) with a structured prompt containing role assignment, node capacities and fidelities, predicted user areas, inference memory history, adjacency constraints, and contrastive few-shot examples, then validates outputs against constraints C6 (adjacency) and C17 (capacity); (3) SET applies Proximity-Greedy Allocation for pre/post-processing function placement and Latency-Biased Shortest Path for routing, with spillover replication when single-node capacity is exceeded.

Load-bearing premise

The LEAD phase assumes that a prompted LLM can produce effective UAV trajectory and inference assignment decisions when given structured prompts with domain rules and few-shot examples. If the LLM produces poor or infeasible assignments, the downstream heuristic phase cannot recover the lost quality. The paper validates LLM outputs against feasibility constraints but does not isolate how much of HyPE's performance depends on the LLM versus the prediction and heuristic phases.

What would settle it

An ablation study replacing the LEAD phase with a random or rule-based trajectory and assignment generator while keeping MAP and SET fixed. If HyPE's coverage and fidelity metrics remain similar, the LLM is not the load-bearing component. Additionally, testing with LLMs of varying capability levels to determine whether decision quality scales with model reasoning ability or plateaus.

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

If this is right

  • If LLM-guided orchestration can approach optimal solutions for NP-hard network problems at a fraction of the computational cost, it could enable real-time adaptive service provisioning in scenarios where traditional optimization solvers cannot respond within the required time window.
  • The contrastive few-shot learning mechanism, where high-reward and low-reward past decisions are fed back into prompts, suggests a pathway for network orchestration systems that improve from operational experience without model retraining.
  • The controlled quality degradation under load, where HyPE shifts inference to distilled models on UAVs to preserve latency and personalization at the cost of fidelity, demonstrates a practical elasticity mechanism for 6G service tiers.
  • The modular three-stage decomposition, prediction-decision-heuristic, could be applied to other combinatorial infrastructure problems where full joint optimization is intractable but partial solutions from different methods can be composed.

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 / 9 minor

Summary. This paper introduces HyPE, a three-phase framework (MAP, LEAD, SET) for personalized AI service provisioning in UAV-assisted 6G networks. The framework combines DRL-based mobility prediction, LLM-guided UAV trajectory and inference assignment decisions, and heuristic-based service placement and routing. The authors formulate a joint optimization problem with 18 constraints covering trajectory planning, service placement, and routing, and propose HyPE as a scalable alternative to the NP-hard optimal solution. Simulations use empirical mobility traces and heterogeneous AI workloads, comparing HyPE against optimization-based, state-of-the-art (AD-SAC, JAAPD-D), and random baselines. The paper targets the IEEE FINE 2026 workshop.

Significance. The integration of LLM-based reasoning for real-time network orchestration decisions in UAV-assisted 6G networks is a novel and timely contribution. The optimization formulation (Section III) is comprehensive, coupling trajectory planning, service placement, and routing under 18 constraints. The use of structured prompting with role assignment, contrastive few-shot learning, and output schema enforcement for the LLM is well-motivated and practically grounded. The simulation setup uses real-world mobility traces (Zenodo) and heterogeneous AI workloads (MMMU-Pro), lending credibility to the evaluation. The polynomial per-frame complexity claim and the feasibility validation pipeline (C6, C17 checks) are commendable. However, the central quantitative claims in the conclusion require better traceability to the experimental results, and the lack of an ablation isolating the LLM contribution is a gap.

major comments (4)
  1. Section VI (Conclusion): The claim that HyPE achieves '91% of the oracle' coverage is not clearly derivable from the reported results. Figure 3 shows HyPE matching the oracle at 5 users (100/100) but achieving 61% vs 75% under heavy load (~81% ratio). The paper never explicitly states how the 91% figure is computed—whether it is an average across all user counts, a weighted measure, or derived from another comparison. This is load-bearing because the 'near-optimal' framing in the abstract and conclusion rests on this number. The authors should either provide the explicit computation or revise the claim to match the reported data.
  2. Section VI (Conclusion) and Section V (Fig. 2b discussion): The claim of '94% of optimal output fidelity' appears inconsistent with the reported '45% fidelity decrease' under load. The paper states that fidelity decreases by 45% as tasks shift to UAVs (Section V), yet the conclusion claims 94% of optimal fidelity. No explicit HyPE-to-oracle fidelity ratio is reported at any operating point. If the oracle also degrades under load, the ratio could still be high, but this is not shown. The authors must either provide the explicit fidelity ratio computation or revise the claim.
  3. Section IV (LEAD) and Section V: There is no ablation isolating the LLM's (LEAD) contribution to HyPE's overall performance. The central claim depends on the LLM producing effective trajectory and inference assignment decisions, but the evaluation does not separate LLM decision quality from DRL prediction quality (MAP) or heuristic placement quality (SET). An ablation replacing LEAD with a simpler decision module (e.g., greedy or random assignment) while keeping MAP and SET fixed would demonstrate the LLM's marginal contribution and substantiate the claim that LLM-based reasoning is superior to simpler decision modules.
  4. Section IV, Algorithm 1, and Section V: The paper does not report LLM API latency or failure rates, which is critical for the real-time scalability claim. The conclusion states polynomial per-frame complexity O(T(N*U + N*U*F + U*N^2 log N)), but this does not account for the LLM API call in step 3. Given that Gemini-2.5-Flash is queried once per frame, the end-to-end latency including API round-trip time could be significant under heavy load. The authors should report measured LLM response times and discuss whether the real-time claim holds, particularly under heavy load.
minor comments (9)
  1. Section III, Constraint C3: The summation on the right-hand side uses index R, which appears to be the request set, but the constraint is quantified over r in R. The notation is confusing—should the sum be over N (nodes) rather than R? Please clarify.
  2. Section III, Constraint C4: The right-hand side has a fraction with P_R X^phi_{r,f,t} in the numerator and R in the denominator, but the quantification uses R in a non-standard way. The constraint should be cleaned up.
  3. Section III, OF: The objective function adds three terms with different units (request count, fidelity score, latency in ms) weighted by alpha, beta, zeta. The paper acknowledges 'different numerical scales' but does not discuss normalization. A brief note on how weights were tuned or whether normalization was applied would improve reproducibility.
  4. Section V, Fig. 2a: The latency figure shows min/max/median with means, but the x-axis labels and user count corresponding to each box are not clearly indicated. Adding explicit user counts or load levels on the x-axis would be clearer.
  5. Section V: The personalization metric values are reported as raw numbers (e.g., '26 vs 12') without units or scale. Clarifying the range or normalization of the personalization score (Q^pe_u in C16) would help readers interpret the magnitude of differences.
  6. Section IV, complexity: The stated complexity O(T(N*U + N*U*F + U*N^2 log N)) should specify which variables correspond to which quantities (e.g., N = nodes, U = users, F = functions). A brief annotation would strengthen the claim.
  7. Table I: The entry for 'Node quality (Q_n)' uses range 0.6-1, but the text refers to fidelity scores without specifying which nodes get which values. A brief note on the mapping between node types (cloud/edge/UAV) and fidelity scores would be helpful.
  8. Section II-C: The personalization quality Q^pe_u is defined as a 'continuity proxy' but the exact interpretation (higher is better? what is the theoretical maximum?) is not stated. A brief clarification in the definition would be valuable.
  9. Figure 1: The prompt template shows '<Requests>' in the Users block but the closing tag is '</Requests>' (plural). Minor inconsistency.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. The referee correctly identifies that the central quantitative claims in the conclusion lack explicit traceability to the experimental results, and that an ablation isolating the LLM's contribution is missing. We address each comment below and commit to revisions where the manuscript needs to change.

read point-by-point responses

  1. Referee: Section VI (Conclusion): The claim that HyPE achieves '91% of the oracle' coverage is not clearly derivable from the reported results. Figure 3 shows HyPE matching the oracle at 5 users (100/100) but achieving 61% vs 75% under heavy load (~81% ratio). The paper never explicitly states how the 91% figure is computed—whether it is an average across all user counts, a weighted measure, or derived from another comparison. This is load-bearing because the 'near-optimal' framing in the abstract and conclusion rests on this number. The authors should either provide the explicit computation or revise the claim to match the reported data.

    Authors: The referee is correct. The 91% figure was computed as the unweighted average of HyPE-to-oracle coverage ratios across all tested user counts (5 through 30 users), but this computation method was never stated in the paper, and the resulting single number obscures the important fact that the ratio degrades substantially under heavy load (from 100% at 5 users to approximately 81% at 30 users). Presenting a single averaged figure without context is misleading given that the ratio varies significantly across operating points. We will revise the conclusion to replace the '91% of the oracle' claim with an explicit description of HyPE's coverage performance across the tested range: matching the oracle at light load (5 users) and maintaining approximately 81% of oracle coverage under heavy load (30 users). We will also add a sentence in Section V explicitly stating how any aggregate ratio is computed, if one is retained. The abstract will be revised to replace 'near-optimal coverage' with 'competitive coverage' or similar language that accurately reflects the reported data. revision: yes

  2. Referee: Section VI (Conclusion) and Section V (Fig. 2b discussion): The claim of '94% of optimal output fidelity' appears inconsistent with the reported '45% fidelity decrease' under load. The paper states that fidelity decreases by 45% as tasks shift to UAVs (Section V), yet the conclusion claims 94% of optimal fidelity. No explicit HyPE-to-oracle fidelity ratio is reported at any operating point. If the oracle also degrades under load, the ratio could still be high, but this is not shown. The authors must either provide the explicit fidelity ratio computation or revise the claim.

    Authors: The referee correctly identifies an inconsistency. The 94% figure was intended to represent the HyPE-to-oracle fidelity ratio averaged across operating points, but this is never explicitly stated, and the referee is right that it appears to conflict with the 45% fidelity decrease reported in Section V. The 45% figure refers to the absolute fidelity decrease of HyPE itself as load increases (due to shifting inference to lower-fidelity UAV-hosted distilled models), while the 94% was meant to capture the ratio relative to the oracle, which also degrades under load. However, since the oracle's fidelity degradation under load is not explicitly reported, the reader cannot verify the ratio. We will revise the conclusion to remove the '94% of optimal output fidelity' claim and replace it with a factual description: HyPE maintains competitive output fidelity relative to the oracle across operating points, with fidelity decreasing by up to 45% under heavy load as a deliberate trade-off to preserve latency and personalization continuity. We will also add the oracle's fidelity values at representative operating points in Section V so that the comparison is transparent. revision: yes

  3. Referee: Section IV (LEAD) and Section V: There is no ablation isolating the LLM's (LEAD) contribution to HyPE's overall performance. The central claim depends on the LLM producing effective trajectory and inference assignment decisions, but the evaluation does not separate LLM decision quality from DRL prediction quality (MAP) or heuristic placement quality (SET). An ablation replacing LEAD with a simpler decision module (e.g., greedy or random assignment) while keeping MAP and SET fixed would demonstrate the LLM's marginal contribution and substantiate the claim that LLM-based reasoning is superior to simpler decision modules.

    Authors: This is a fair and important point. The current evaluation compares HyPE (as a complete framework) against external baselines, but does not isolate the marginal contribution of the LEAD phase's LLM-based reasoning versus simpler decision modules for trajectory and inference assignment. We agree that such an ablation is necessary to substantiate the claim that LLM-based reasoning provides value beyond what MAP (DRL prediction) and SET (heuristics) contribute. We will add an ablation study in which LEAD is replaced by a greedy trajectory-and-assignment module (selecting UAV moves and inference nodes based on proximity and capacity, without LLM reasoning) while keeping MAP and SET unchanged. This will be included as an additional configuration in Section V. We expect this to show that the LLM's contextual reasoning over personalization history, fidelity requirements, and mobility predictions contributes measurably to HyPE's performance, but we will report the results honestly regardless of the outcome. revision: yes

  4. Referee: Section IV, Algorithm 1, and Section V: The paper does not report LLM API latency or failure rates, which is critical for the real-time scalability claim. The conclusion states polynomial per-frame complexity O(T(N*U + N*U*F + U*N^2 log N)), but this does not account for the LLM API call in step 3. Given that Gemini-2.5-Flash is queried once per frame, the end-to-end latency including API round-trip time could be significant under heavy load. The authors should report measured LLM response times and discuss whether the real-time claim holds, particularly under heavy load.

    Authors: The referee raises a valid concern. The polynomial complexity analysis in the conclusion covers only the computational steps in MAP, SET, and the validation pipeline, and does not account for the LLM API round-trip latency in LEAD (step 3 of Algorithm 1). Since Gemini-2.5-Flash is queried once per frame, the API response time is a non-negligible component of the per-frame decision latency and should be reported. We will add measured LLM API response times (mean and percentiles) to Section V, along with the rate of infeasible outputs that require re-querying (failure/retry rate). We will also add a discussion of whether the end-to-end per-frame latency, including the API call, remains within the real-time budget implied by the frame duration. If the API latency is significant relative to the frame duration under heavy load, we will state this explicitly and discuss implications for deployment, including potential mitigations such as local model deployment or asynchronous querying. The complexity expression in the conclusion will be annotated to note that it excludes the LLM API call, which is bounded by the external service response time rather than by problem size. revision: yes

Circularity Check

0 steps flagged

No significant circularity; self-citations are methodological building blocks, not circular validation of results

full rationale

The paper's derivation chain is largely self-contained. The three HyPE phases (MAP→LEAD→SET) are independently motivated: MAP trains a DRL predictor on prediction accuracy (not downstream task metrics), LEAD uses an LLM with structured prompts and contrastive few-shot examples, and SET applies standard greedy heuristics. The self-citations ([17] for NP-hardness, [18] for DRL prediction methods, [21] for LLM orchestration) are by the same authors but serve as methodological references, not as circular validation of HyPE's claims. NP-hardness of joint trajectory/service placement/routing is a standard result. The DRL predictor is trained on prediction accuracy independently of the downstream optimization objective. The contrastive few-shot examples in LEAD use past assignment rewards (latency, fidelity, personalization scores) as training signal—this is standard online learning, not circularity. The framework is evaluated against external baselines (AD-SAC [15], JAAPD-D [16], Gurobi optimization, random) using real-world datasets (Zenodo mobility traces, MMMU-Pro). The headline metrics in the conclusion (91% coverage, 94% fidelity) may be insufficiently supported by the experimental data as reported (e.g., the 45% fidelity decrease under load vs. the 94% claim), but this is a correctness/supportability concern, not circularity—the numbers are not derived from self-referential definitions or fitted parameters renamed as predictions. No variable or claim reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 1 invented entities

See above.

free parameters (7)
  • α (fidelity weight) = not specified
    Objective function weight for output fidelity term; tuned to balance relative importance of fidelity, personalization, and latency (Section III).
  • β (personalization weight) = not specified
    Objective function weight for personalization quality term (Section III).
  • ζ (latency penalty) = not specified
    Objective function weight for latency penalty term (Section III).
  • δ (personalization decay factor) = 0.65–0.9
    Controls recency weighting in personalization history H_pa_{u,n,t} (C15); range given in Table I.
  • κ (binding weight) = 0.1–1.0
    Balances recency vs latency in user-node binding (Eq. 2); range given in Table I.
  • γ (congestion avoidance bias) = 0.1–1.0
    Tunes congestion avoidance in LBSP routing weight (Eq. 5); range given in Table I.
  • K (few-shot examples) = 1-3
    Number of positive/negative examples in LLM contrastive prompting (Table I).
axioms (4)
  • domain assumption The joint trajectory planning, service placement, and routing problem is NP-hard
    Stated in Section IV opening paragraph, cited to [17] (self-citation). Used to justify the need for heuristic alternatives.
  • domain assumption DRL agents can predict user mobility and service request distributions with sufficient accuracy to guide proactive provisioning
    Underlies the MAP phase (Section IV). The paper claims online DRL captures spatio-temporal evolution but does not provide prediction accuracy metrics in the evaluation.
  • ad hoc to paper LLMs can produce effective network orchestration decisions when prompted with domain rules and few-shot examples
    Underlies the LEAD phase (Section IV). No formal guarantee or lower bound on LLM decision quality is provided. The evaluation does not isolate LLM contribution via ablation.
  • domain assumption Personalization quality can be proxied by recency-weighted history of user-node inference assignments
    Constraint C15-C16 defines Q_pe_u as sum of decay-weighted historical affinities. This is a modeling choice; whether it captures real personalization quality is assumed, not validated.
invented entities (1)
  • QoAIS (Quality of AI Services) as formalized in the paper no independent evidence
    purpose: Captures output fidelity, continuity, and personalization as a unified quality metric
    The specific formalization of QoAIS via Q_ou_r (fidelity) and Q_pe_u (personalization continuity) is introduced in this paper. The personalization proxy (recency-weighted history) is not validated against user-perceived quality metrics.

pith-pipeline@v1.1.0-glm · 15902 in / 4607 out tokens · 330406 ms · 2026-07-08T11:00:15.773155+00:00 · methodology

0 comments
read the original abstract

In sixth-generation (6G) artificial intelligence (AI) services, two quality dimensions should be jointly addressed: conventional quality (e.g., latency) and Quality of AI Services (QoAIS; output fidelity, continuity, personalization). Existing methods emphasize conventional quality, while neglecting QoAIS, particularly for personalized outputs in dynamic aerial-terrestrial settings. This paper introduces HyPE, a Hybrid Predictive-in-context-lEarning framework for holistically quality-aware personalized AI service provisioning in Unmanned Aerial Vehicle (UAV)-assisted 6G networks. HyPE integrates: (i) mobility-aware prediction to forecast spatio-temporal request distributions, (ii) learning-augmented decision leveraging Large Language Model (LLM)-based reasoning to optimize UAV trajectories and inference assignments, and (iii) pre-/post-processing service placement and routing using heuristics. We formulate an optimization problem for joint trajectory planning, service placement, and routing, and present HyPE as a scalable alternative to intractable optimal solutions. Simulations with empirical mobility traces and heterogeneous AI workloads show near-optimal coverage, reduced end-to-end latency, sustained QoAIS-driven, and continuity-based service personalization versus optimization and state-of-the-art baselines. The results highlight the promise of predictive learning-augmented provisioning for elastic, user-centric AI in 6G.

Figures

Figures reproduced from arXiv: 2607.06278 by Masoud Shokrnezhad, Mohammad Farhoudi, Tarik Taleb.

Figure 1
Figure 1. Figure 1: Structured prompt design for LLM-based node trajectory and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: b evaluates AI-enabled personalization (continuity [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of HyPE with Optimize (OF), AD-SAC [15], JAAPD-D [16], and random methods in terms of a) latency, and b) AI-enabled capabilities [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coverage comparison in terms of accepted requests as the number [PITH_FULL_IMAGE:figures/full_fig_p008_3.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

25 extracted references · 25 canonical work pages · 3 internal anchors

  1. [1]

    Service registration, indexing, discovery, and selection: An architectural survey toward a GenAI-driven future,

    M. Farhoudi, M. Shokrnezhad, and T. Taleb, “Service registration, indexing, discovery, and selection: An architectural survey toward a GenAI-driven future,”IEEE Access, vol. 13, pp. 209 680–209 722, 2025

    work page 2025

  2. [2]

    Adaptive multiple access and service placement for generative diffusion models,

    H. Mazandaraniet al., “Adaptive multiple access and service placement for generative diffusion models,” inProc. IEEE Global Telecommun. Conf., Taipei, Taiwan, 2025, p. 5.97

    work page 2025

  3. [3]

    A survey of self-evolving agents: On path to artificial super intelligence,

    H. ang Gaoet al., “A survey of self-evolving agents: On path to artificial super intelligence,” 2025

    work page 2025

  4. [4]

    Optimization of quality of AI service in 6G native AI wireless networks,

    T. Chenet al., “Optimization of quality of AI service in 6G native AI wireless networks,”Electronics, vol. 12, no. 15, 2023

    work page 2023

  5. [5]

    A survey of graph retrieval-augmented generation for customized large language models,

    Q. Zhanget al., “A survey of graph retrieval-augmented generation for customized large language models,” 2025

    work page 2025

  6. [6]

    A Survey on Collaborative Mechanisms Between Large and Small Language Models

    Y . Chen, J. Zhao, and H. Han, “A survey on collaborative mechanisms between large and small language models,” 2025. [Online]. Available: https://arxiv.org/abs/2505.07460

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Rap: Retrieval-augmented personalization for multimodal large language models,

    H. Hao, J. Han, C. Li, Y .-F. Li, X. Yueet al., “Rap: Retrieval-augmented personalization for multimodal large language models,” 2025

    work page 2025

  8. [8]

    Personalized multimodal large language models: A survey,

    J. Wuet al., “Personalized multimodal large language models: A survey,” 2024

    work page 2024

  9. [9]

    Discovery of 6G services and resources in edge-cloud-continuum,

    M. Farhoudi, M. Shokrnezhadet al., “Discovery of 6G services and resources in edge-cloud-continuum,”IEEE Netw., vol. 39, no. 3, pp. 223–232, 2025

    work page 2025

  10. [10]

    6G system architecture: A service of services vision,

    T. Talebet al., “6G system architecture: A service of services vision,” in ITU journal on future and evolving technologies, vol. 3, no. 3, pp. 710–743, Dec. 2022

    work page 2022

  11. [11]

    Semantic-aware dynamic and distributed power allocation: a multi-UA V area coverage use case,

    H. Mazandaraniet al., “Semantic-aware dynamic and distributed power allocation: a multi-UA V area coverage use case,” 2025

    work page 2025

  12. [12]

    Joint task offloading, resource allocation, and trajectory design for multi-UA V cooperative edge computing with task priority,

    H. Hao, C. Xu, W. Zhang, S. Yang, and G.-M. Muntean, “Joint task offloading, resource allocation, and trajectory design for multi-UA V cooperative edge computing with task priority,”IEEE Trans. Mobile Comput., vol. 23, no. 9, pp. 8649–8663, 2024

    work page 2024

  13. [13]

    Adaptive heuristics for scheduling DNN inferencing on edge and cloud for personalized UA V fleets,

    S. Raj, R. Mittal, H. Gupta, S. Yogesh SimmhanRajet al., “Adaptive heuristics for scheduling DNN inferencing on edge and cloud for personalized UA V fleets,”Future Generation Computer Systems, vol. 173, p. 107874, Dec 2025

    work page 2025

  14. [14]

    Adaptive task offloading and resource management for vehicular edge computing,

    L. Jinet al., “Adaptive task offloading and resource management for vehicular edge computing,”IEEE Trans. Veh. Technol., pp. 1–14, 2025

    work page 2025

  15. [15]

    Accuracy-aware MLLM task offloading and resource allocation in UA V-assisted satellite edge computing,

    H. Yanet al., “Accuracy-aware MLLM task offloading and resource allocation in UA V-assisted satellite edge computing,”Drones, vol. 9, no. 7, 2025

    work page 2025

  16. [16]

    AI service deployment and resource allocation optimization based on human-like networking architecture,

    Y . Huet al., “AI service deployment and resource allocation optimization based on human-like networking architecture,”IEEE In- ternet Things J., vol. 11, no. 14, pp. 24 795–24 813, 2024

    work page 2024

  17. [17]

    Deep learn- ing based service composition in integrated aerial-terrestrial networks,

    M. Farhoudi, M. Shokrnezhad, S. Kianpisheh, and T. Taleb, “Deep learn- ing based service composition in integrated aerial-terrestrial networks,” inInternational Conf. on Net. Softwarization, 2025, pp. 204–208

    work page 2025

  18. [18]

    QoS-aware service prediction and orchestration in cloud-network integrated beyond 5G,

    M. Farhoudiet al., “QoS-aware service prediction and orchestration in cloud-network integrated beyond 5G,” inProc. IEEE Global Telecom- mun. Conf., Dec. 2023, pp. 369–374

    work page 2023

  19. [19]

    From unstructured data to in-context learning: Exploring what tasks can be learned and when,

    K. C. Wibisono and Y . Wang, “From unstructured data to in-context learning: Exploring what tasks can be learned and when,” inAdvances in Neural Information Processing Systems, 2024, pp. 16 369–16 405

    work page 2024

  20. [20]

    Better Zero-Shot Reasoning with Role-Play Prompting

    A. Konget al., “Better zero-shot reasoning with role-play prompting,” arXiv preprint arXiv:2308.07702, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  21. [21]

    An autonomous network orchestration frame- work integrating large language models with continual reinforcement learning,

    M. Shokrnezhadet al., “An autonomous network orchestration frame- work integrating large language models with continual reinforcement learning,”IEEE Commun. Mag., vol. 63, no. 8, pp. 78–84, 2025

    work page 2025

  22. [22]

    Pydantic: Data validation using python type hints,

    S. Colvinet al., “Pydantic: Data validation using python type hints,” MIT License. [Online]. Available: https://github.com/pydantic/pydantic

    work page

  23. [23]

    Metropolitan Scale and Longitudinal Dataset of Anonymized Human Mobility Trajectories

    T. Yabeet al., “Metropolitan scale and longitudinal dataset of anonymized human mobility trajectories,” 2023. [Online]. Available: https://arxiv.org/abs/2307.03401

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [24]

    Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark,

    X. Yueet al., “Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark,” 2025

    work page 2025

  25. [25]

    M.2160-0: Framework and overall objectives of the future development of IMT for 2030 and beyond,

    R. I. R. Sector, “M.2160-0: Framework and overall objectives of the future development of IMT for 2030 and beyond,” Nov. 2023, available: https://www.itu.int/rec/RREC-M.2160/en

    work page 2030