Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

HybridRAG-based LLM Agents for Low-Carbon Optimization in Low-Altitude Economy Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper argues that fusing keyword, vector, and graph retrieval lets LLM agents formulate more accurate carbon-minimization problems for multi-UAV edge networks, and that a double-regularized diffusion reinforcement learning algorithm…

desk verdict The R2DSAC side is a coherent, fairly complete engineering effort, but the central HybridRAG claim rests on a test set generated by the system under test, so the headline claim does not survive. read the letter →

arxiv 2506.15947 v1 pith:7ARGE7GZ submitted 2025-06-19 cs.NI eess.SP

Jinbo Wen , Cheng Su , Jiawen Kang , Jiangtian Nie , Yang Zhang , Jianhang Tang , Dusit Niyato , Chau Yuen This is my paper · ORCID
classification cs.NIeess.SP
keywords low-altitudeeconomynetworksmulti-UAVmobileedgecomputingretrievalaugmentedgenerationgraph-basedRAGLLMagentdiffusionreinforcementlearningcarbonemissionoptimizationdynamicneuronpruning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that an LLM agent can take over the delicate job of formulating a carbon-minimization optimization problem for multi-UAV mobile edge computing networks, and that the problem it writes can be solved efficiently. To do this, it fuses three retrieval strategies — keyword, vector, and graph — into a single HybridRAG module that pulls governing equations and relational structures from expert documents. It then proposes a diffusion-based reinforcement learning algorithm, R²DSAC, with two entropy regularizations and dynamic neuron pruning, to solve the formulated mixed-integer NP-hard problem. If correct, the framework reduces human modeling errors and produces greener network operation strategies, with a reported 64% improvement over soft actor-critic in the test scenario.

What carries the argument

HybridRAG: a retrieval module that merges three channels — KeywordRAG (keyword matching over hierarchical document indices), VectorRAG (semantic similarity over embedded chunks), and GraphRAG (triplet-structured knowledge graph stored in Neo4j, built by LLM-based triplet extraction from expert documents). The final retrieval $C_{\mathrm{final}} = C_{\mathrm{keyword}} \cup C_{\mathrm{graph}} \cup C_{\mathrm{vector}}$ is fed with the user query to the LLM agent. The solver is R²DSAC, a soft actor-critic variant whose policy is a conditional diffusion model: actions are generated by reverse denoising from Gaussian noise, with Q-learning guidance, and the policy objective combines action entropy regularization (encouraging exploration) and diffusion entropy regularization (behavior-cloning toward high-value actions). A dynamic pruning module masks the least important neurons in the actor network according to a pruning rate $\varrho$ to reduce training carbon emissions.

What would settle it

Run the same RAGChecker evaluation on an independently authored set of optimization-formulation QA pairs (for example, drawn from published multi-UAV MEC papers with known answers) and check whether HybridRAG's F1, CR, and CU gains over VectorRAG+KeywordRAG persist; likewise, solve the problems formulated by HybridRAG and by a human expert with the same R²DSAC solver and compare the achieved carbon emissions under identical simulation settings.

Watch

Extended reading notes

Core claim

The central claim is that combining KeywordRAG, VectorRAG, and GraphRAG into a single retrieval pipeline enables LLM agents to formulate carbon emission optimization problems for multi-UAV-assisted MEC networks more accurately than traditional RAG does, because the graph component supplies relational structure (e.g., LoS probability affecting G2A links) that semantic similarity alone misses. The paper further claims that the formulated problem — minimizing weighted carbon emissions through joint task offloading, computing resource allocation, and UAV trajectory control — is a non-convex NP-hard mixed-integer program, and that the proposed R²DSAC algorithm solves it effectively. R²DSAC uses a diffusion policy with diffusion entropy regularization and action entropy regularization to stabilize learning, plus dynamic pruning of unimportant neurons in the actor network to lower the carbon cost of training; simulations report the lowest carbon emissions among compared baselines and a 64% test-reward gain over SAC.

Load-bearing premise

The claim that HybridRAG produces more accurate optimization problems than traditional RAG rests on a test dataset of question-answer pairs that the HybridRAG-based LLM agent itself generated; if those self-made pairs favor the structures HybridRAG retrieves, the improved metrics would not demonstrate real accuracy gains.

Editorial extensions

If this is right

  • Network designers can interactively formulate a carbon-minimization problem for UAV-MEC networks in natural language, and the HybridRAG agent supplies the governing equations and constraints from expert documents, reducing the risk of human errors like omitting flight propulsion energy.
  • HybridRAG outperforms combined VectorRAG+KeywordRAG on claim-level F1 (53.2 vs 49.9), claim recall (83.1 vs 82.7), and context utilization (80.2 vs 75.8) in the paper's evaluation, implying that adding the graph retrieval channel improves both retrieval and generation quality for these structured network optimization questions.
  • R²DSAC achieves the lowest carbon emissions among compared algorithms and a 64% gain over SAC in test rewards, with dynamic pruning keeping training carbon emissions at about 70.3 g total and about 0.025 g per inference.
  • The framework is portable: because the retrieval database is external, the same HybridRAG agent can be retargeted to other network optimization tasks by swapping the expert documents.

Reading between the lines

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

  • Editorial inference: the RAG comparison is tested on QA pairs generated by the HybridRAG agent itself (Fig. 5); a fairer test would use independently expert-curated formulation problems, because self-generated questions may embed the same structural biases the graph channel is designed to satisfy.
  • Editorial inference: the 64% gain over SAC and the pruning results are demonstrated in a single simulated scenario (2 UAVs, 10 users, 1000 m × 1000 m); scaling to larger networks and denser user distributions would test whether the diffusion policy's advantage persists.
  • Testable extension: the paper does not report carbon emissions of R²DSAC against SAC with matched training budgets; measuring emissions at equal final-reward levels would separate the pruning benefit from the policy's performance benefit.
  • Testable extension: since the knowledge graph construction is LLM-driven (triplet extraction), one could automatically update the graph when new expert papers are added, making the formulation process continuously self-improving.
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 HybridRAG-based LLM agents that combine KeywordRAG, VectorRAG, and GraphRAG to help formulate carbon-emission minimization problems for multi-UAV mobile-edge-computing networks, and a diffusion-enhanced Soft Actor-Critic algorithm (R2DSAC) with double entropy regularization and dynamic neuron pruning to solve the formulated problem. The authors claim that HybridRAG generates more accurate optimization problems than traditional RAG approaches and that R2DSAC achieves the lowest carbon emissions with a 64% gain over SAC. The paper includes a full system model, the proposed problem formulation, an MDP reformulation, algorithm details, complexity analysis, and simulation results using RAGChecker and DRL benchmarks.

Significance. If the claims were valid, the framework would offer a semi-automated pipeline for low-carbon UAV-MEC optimization, with a plausible mechanism for reducing human modeling errors and a sustainability-aware training method through pruning. The paper is timely given the interest in LLM-based network optimization, and the incorporation of graph-structured retrieval is a meaningful direction. However, the experimental validation of the central HybridRAG superiority claim is methodologically flawed, and the DRL evaluation lacks statistical rigor. As a result, the significance of the contribution is not currently established.

major comments (4)
  1. [V.B, Fig. 5] The baseline test dataset for evaluating HybridRAG is built from QA pairs generated by the HybridRAG-based LLM agent itself. This makes the evaluation circular: the system under test authors both the questions and the ground-truth answers, so any retrieval pipeline that returns the same context is rewarded, while the VectorRAG+KeywordRAG baseline is systematically disadvantaged. The central claim that HybridRAG generates "more accurate optimization problems" is therefore unsupported. The paper must evaluate on independently constructed QA pairs (e.g., expert-written or manually validated) before this claim can be accepted.
  2. [V.B, Table II] Even within the circular benchmark, HybridRAG performs worse than VectorRAG+KeywordRAG on Hallucination (7.2 vs. 6.7) and Faithfulness (92.4 vs. 92.8). The text acknowledges these deficits but does not explain why they are acceptable; hallucinations are especially problematic for optimization problem formulation because they can yield incorrect constraints or objectives. The overall superiority claim is not uniformly supported by the data in Table II.
  3. [V.C, Fig. 6] The R2DSAC comparisons against SAC, PPO, random, and the BCDSAC/TDSAC/DSAC ablations are presented as single learning curves with no error bars, no multiple seeds, and no significance testing. The claimed "64% gain" and the ablation ordering could be due to random seed effects or hyperparameter tuning on the same simulation. The paper should report mean plus/minus standard deviation over several independent runs, and the sensitivity analysis in Fig. 7 should use a held-out environment or a proper model-selection criterion.
  4. [V.B, Table II] The HybridRAG evaluation compares only against VectorRAG+KeywordRAG; there is no comparison to GraphRAG alone or to other retrieval combinations. Since GraphRAG is the main novel component of the proposed HybridRAG, the specific contribution of the graph module is not isolated, and the paper does not demonstrate that the graph component adds value beyond the fusion of keyword and vector retrieval.
minor comments (5)
  1. [IV.B, Eq. (37)] The TD target uses Q_\hat{\phi}(s(n+1)) without an action argument, which is inconsistent with the Q-function definition Q_\phi(s,a); please clarify the notation or correct the equation.
  2. [V.C] The text says "R2DSAC achieves a 64% performance improvement over SAC," while the figure caption refers to a "64.17% gain" in average carbon emissions; please unify the terminology to avoid ambiguity.
  3. [III.C, Theorem 1] The proof of NP-hardness asserts NP-hardness from the presence of binary variables without a reduction from a known NP-hard problem; either provide a proper reduction or soften the claim to "mixed-integer nonlinear programming," which is generally intractable.
  4. [V.B, Fig. 5] The JSON example in Fig. 5 contains malformed quoting (the "standalone_eval" field appears with a curly quote) and does not explain the scale or meaning of groundedness_score and relevance_score; please correct and describe these scoring scales.
  5. [Throughout] The abbreviation "UA Vs" is inconsistently spaced, and "LAENets" is used in the abstract without a defined expansion; please fix these presentation issues.

Circularity Check

1 steps flagged · score 6.0 of 10

HybridRAG evaluation is circular: the QA test set is generated by the HybridRAG agent itself, so Table II cannot support the claimed superiority; R2DSAC is externally benchmarked.

  1. fitted input called prediction [Section V.B (Performance Evaluation of HybridRAG), Fig. 5]
    "Prior to evaluation, we construct a baseline test dataset consisting of Question-Answer (QA) pairs generated by the HybridRAG-based LLM agent, as shown in Fig. 5. ... Fig. 5: QA pairs outputted by the HybridRAG-based LLM agent, which can serve as the test dataset to evaluate the performance of HybridRAG."

    The ground-truth QA pairs used to score retrieval and generation are produced by the same HybridRAG system whose accuracy the evaluation claims to test. RAGChecker's claim-level recall/precision, CR, and CU metrics compare retrieved chunks and generated claims against these self-authored answers. Because the answers were generated with HybridRAG's own retrieval context, the hybrid retriever is rewarded for returning that same context, while the VectorRAG+KeywordRAG baseline is systematically disadvantaged. The Table II gains (F1 49.9->53.2, CR 82.7->83.1, CU 75.8->80.2) therefore reflect self-consistency, not demonstrated superiority. Notably Hallu. and Faith. move in the opposite direction (7.2 vs 6.7 and 92.4 vs 92.8), further weakening the claim.

full rationale

The paper's central HybridRAG claim is evaluated against a test dataset generated by HybridRAG itself, so the headline comparison to VectorRAG+KeywordRAG is circular. This is the only load-bearing circular step found. The R2DSAC algorithm is tested against external DRL baselines (SAC, PPO) and its ablation variants, so that contribution has independent empirical content. Equations in Sections III-IV are standard model definitions and diffusion-policy losses from cited prior work, not reduced to the paper's own outputs. The problem formulation is generated by the LLM agent and then solved by R2DSAC, but that is a normal pipeline, not circularity. Thus score is 6: partial circularity in the central RAG evaluation, with the RL contribution independently benchmarked.

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

No new physical entities, forces, dimensions, or particles are introduced; HybridRAG and R2DSAC are software systems built from existing components. The main pulls from prior literature are the standard UAV mobility, communication, energy, and propulsion models, plus the diffusion policy and SAC machinery.

free parameters (5)
  • Pruning rate ϱ = 0.1 (swept 0.1-0.9)
    Chosen from Fig. 7(a) on the same simulation used for final performance; determines how many neurons are masked and thus claimed training-carbon savings.
  • Behavior-cloning weight ρ = not reported (swept in [0,1])
    Weights action-entropy and diffusion cloning losses in Eq. (40); no optimal value stated, and the sweep in Fig. 7(c) is on the test simulation.
  • Diffusion steps T = 3 (swept in Fig. 7(b))
    Number of reverse denoising steps; selected on the same simulation, directly affecting training time and carbon emission estimates.
  • Reward-shaping penalties Ω_d, Ω_f, Ω_g, Ω_i = not reported
    Hand-designed constraint terms in Eq. (30) using "human knowledge" [34]; their magnitudes are not given, yet they shape the learned policy and reported rewards.
  • DRL and diffusion hyperparameters (learning rate, batch size, discount factor, temperature, target update rate, ψ_min… = not reported
    Required to reproduce R2DSAC; absent from Table I and Section V.A.
assumptions (5)
  • domain assumption Carbon emissions are proportional to total energy consumption with a fixed coefficient ς_Carbon and conversion factor τ (Eq. 18).
    Invoked in Section III.A.6 from refs [5],[30]; treats drone computation and propulsion as producing CO2 linearly, ignoring lifecycle or energy source variations.
  • domain assumption Probabilistic LoS/NLoS path loss model with fixed environment constants a,b and excess losses η_LoS, η_NLoS (Eqs. 7-10).
    Adopted from refs [7],[8]; if the environment parameters do not hold, transmission rate and energy expressions change.
  • domain assumption Each user offloads each task to exactly one UAV at each slot and downlink cost is ignored (Eqs. 15, 13).
    From refs [3],[5]; simplifies scheduling but may not reflect practical partial offloading or downlink energy.
  • ad hoc to paper The optimization problem in Eq. (27) is NP-hard and non-convex (Theorem 1).
    The proof only notes binary variables and a non-convex constraint (27b); no reduction from a known NP-hard problem is given, so the theorem is an unproved assertion used to justify DRL.
  • domain assumption The reward-shaping terms in Eq. (30) correctly enforce constraints via human knowledge and linear mapping.
    Section IV.A.3 states other constraints can be guaranteed through linear mapping; no detailed construction or validation of Ω terms is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HybridRAG-based LLM Agents for Low-Carbon Optimization in Low-Altitude Economy Networks." pith.science (2026). https://pith.science/paper/7ARGE7GZ

@misc{pith2026250615947,
  author       = {Pith},
  title        = {Pith review of: HybridRAG-based LLM Agents for Low-Carbon Optimization in Low-Altitude Economy Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ARGE7GZ}},
  note         = {Machine review of arXiv:2506.15947}
}
read the original abstract

Low-Altitude Economy Networks (LAENets) are emerging as a promising paradigm to support various low-altitude services through integrated air-ground infrastructure. To satisfy low-latency and high-computation demands, the integration of Unmanned Aerial Vehicles (UAVs) with Mobile Edge Computing (MEC) systems plays a vital role, which offloads computing tasks from terminal devices to nearby UAVs, enabling flexible and resilient service provisions for ground users. To promote the development of LAENets, it is significant to achieve low-carbon multi-UAV-assisted MEC networks. However, several challenges hinder this implementation, including the complexity of multi-dimensional UAV modeling and the difficulty of multi-objective coupled optimization. To this end, this paper proposes a novel Retrieval Augmented Generation (RAG)-based Large Language Model (LLM) agent framework for model formulation. Specifically, we develop HybridRAG by combining KeywordRAG, VectorRAG, and GraphRAG, empowering LLM agents to efficiently retrieve structural information from expert databases and generate more accurate optimization problems compared with traditional RAG-based LLM agents. After customizing carbon emission optimization problems for multi-UAV-assisted MEC networks, we propose a Double Regularization Diffusion-enhanced Soft Actor-Critic (R\textsuperscript{2}DSAC) algorithm to solve the formulated multi-objective optimization problem. The R\textsuperscript{2}DSAC algorithm incorporates diffusion entropy regularization and action entropy regularization to improve the performance of the diffusion policy. Furthermore, we dynamically mask unimportant neurons in the actor network to reduce the carbon emissions associated with model training. Simulation results demonstrate the effectiveness and reliability of the proposed HybridRAG-based LLM agent framework and the R\textsuperscript{2}DSAC algorithm.

Figures

Figures reproduced from arXiv: 2506.15947 by the authors.

Figure 1
Figure 1. An illustration of LAENets, where UAVs emit carbon [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. A structured and queryable knowledge graph for the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The formulation process of a carbon emission optimization problem for multi-UAV-assisted MEC networks by using [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The architecture of the R2DSAC algorithm, with two improvements to traditional diffusion-based DRL algorithms. The first improvement is that we incorporate diffusion en￾tropy regularization and action entropy regularization into the diffusion policy, thereby enhancing …
Figure 5
Figure 5. Figure 5: QA pairs outputted by the HybridRAG-based LLM [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Performance evaluation of the R2DSAC algorithm in carbon emission optimization. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Pruning Rates 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Normalized Values Final rewards Training time Carbon emissions (a) Pruning rate impact. 2 4 6 8 10 …
Figure 7
Figure 7. Figure 7: Impacts of pre-defined parameters on final rewards, training time, and carbon emissions during model training. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: UAV trajectories generated by the R2DSAC algorithm within a single episode under varying environmental states. 0.025 g per inference. In [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

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. TeaRAG: A Token-Efficient Agentic Retrieval-Augmented Generation Framework

    cs.IR 2025-11 conditional novelty 6.0 of 10

    TeaRAG shows that hybrid chunk+triplet retrieval with Personalized PageRank and an iterative process-aware DPO reward keeps QA accuracy while cutting reasoning tokens by roughly 60%.

  2. A Systematic Survey on Large Language Models for Evolutionary Optimization: From Modeling to Solving

    cs.NE 2025-09 conditional novelty 4.0 of 10

    A literature survey that classifies LLM-based optimization research into modeling and solving, with solving divided into LLMs as optimizers, low-level components, and high-level managers.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages · cited by 2 Pith papers

  1. [1]

    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

    arXiv 2025

  2. [2]

    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

    arXiv 2025

  3. [3]

    Multi-objective optimization for multi-UA V-assisted mobile edge computing,

    G. Sun, Y . Wang, Z. Sun, Q. Wu, J. Kang, D. Niyato, and V . C. M. Leung, “Multi-objective optimization for multi-UA V-assisted mobile edge computing,”IEEE Transactions on Mobile Computing, vol. 23, no. 12, pp. 14 803–14 820, 2024

    work page 2024

  4. [4]

    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 Transactions on Mobile Computing, vol. 23, no. 9, pp. 8649–8663, 2024. 14

    work page 2024

  5. [5]

    Efficient and emission-reducing blockchain-enabled multi-UA V-assisted MEC system in IoT networks,

    L. Yu, B. Li, Y . Yao, Z. Wang, Z. Li, Z. Dong, and D. Cai, “Efficient and emission-reducing blockchain-enabled multi-UA V-assisted MEC system in IoT networks,”IEEE Internet of Things Journal, vol. 11, no. 24, pp. 40 645–40 655, 2024

    work page 2024

  6. [6]

    Service experience oriented cooperative computing in cache-enabled UA Vs assisted MEC networks,

    X. Gao and L. Zhai, “Service experience oriented cooperative computing in cache-enabled UA Vs assisted MEC networks,”IEEE Transactions on Mobile Computing, vol. 23, no. 10, pp. 9721–9736, 2024

    work page 2024

  7. [7]

    Multi-UAV Enabled MEC Networks: Optimizing Delay through Intelligent 3D Trajectory Planning and Resource Allocation

    Z. Wang, T. Wei, G. Sun, X. Liu, H. Yu, and D. Niyato, “Multi-UA V enabled MEC networks: Optimizing delay through intelligent 3D trajec- tory planning and resource allocation,”arXiv preprint arXiv:2409.17882, 2024

    work page Pith review arXiv 2024

  8. [8]

    Multi-UA V cooperative task offloading and resource allocation in 5G advanced and beyond,

    H. Guo, Y . Wang, J. Liu, and C. Liu, “Multi-UA V cooperative task offloading and resource allocation in 5G advanced and beyond,”IEEE Transactions on Wireless Communications, vol. 23, no. 1, pp. 347–359, 2024

    work page 2024

Show all 41 references
  1. [9]

    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, pp. 3581–3596, 2024

    2024

  2. [10]

    Multi-agent reinforcement learning with policy clipping and average evaluation for UA V-assisted communication Markov game,

    Z. Feng, M. Huang, D. Wu, E. Q. Wu, and C. Yuen, “Multi-agent reinforcement learning with policy clipping and average evaluation for UA V-assisted communication Markov game,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 12, pp. 14 281–14 293, 2023

    2023

  3. [11]

    Generative AI for low-carbon artificial Intelligence of Things with large language models,

    J. Wen, R. Zhang, D. Niyato, J. Kang, H. Du, Y . Zhang, and Z. Han, “Generative AI for low-carbon artificial Intelligence of Things with large language models,”IEEE Internet of Things Magazine, vol. 8, no. 1, pp. 82–91, 2025

    2025

  4. [12]

    Retrieval- augmented generation for knowledge-intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschelet al., “Retrieval- augmented generation for knowledge-intensive NLP tasks,”Advances in neural information processing systems, vol. 33, pp. 9459–9474, 2020

    2020

  5. [13]

    Retrieval-augmented generation for large language models: A survey,

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, H. Wang, and H. Wang, “Retrieval-augmented generation for large language models: A survey,”arXiv preprint arXiv:2312.10997, vol. 2, no. 1, 2023

    2023 arXiv

  6. [14]

    Graph retrieval-augmented generation: A survey,

    B. Peng, Y . Zhu, Y . Liu, X. Bo, H. Shi, C. Hong, Y . Zhang, and S. Tang, “Graph retrieval-augmented generation: A survey,”arXiv preprint arXiv:2408.08921, 2024

    2024 arXiv

  7. [15]

    When graph meets retrieval augmented generation for wireless networks: A tutorial and case study,

    Y . Xiong, R. Zhang, Y . Liu, D. Niyato, Z. Xiong, Y .-C. Liang, and S. Mao, “When graph meets retrieval augmented generation for wireless networks: A tutorial and case study,”arXiv preprint arXiv:2412.07189, 2024

    2024 arXiv

  8. [16]

    Diffusion-based reinforcement learning for edge-enabled AI-generated content services,

    H. Du, Z. Li, D. Niyato, J. Kang, Z. Xiong, H. Huang, and S. Mao, “Diffusion-based reinforcement learning for edge-enabled AI-generated content services,”IEEE Transactions on Mobile Computing, vol. 23, no. 9, pp. 8902–8918, 2024

    2024

  9. [17]

    From generative AI to generative Internet of Things: Fundamentals, framework, and outlooks,

    J. Wen, J. Nie, J. Kang, D. Niyato, H. Du, Y . Zhang, and M. Guizani, “From generative AI to generative Internet of Things: Fundamentals, framework, and outlooks,”IEEE Internet of Things Magazine, vol. 7, no. 3, pp. 30–37, 2024

    2024

  10. [18]

    Graph-attention-based reinforcement learning for trajectory design and resource assignment in multi-UA V-assisted communication,

    Z. Feng, D. Wu, M. Huang, and C. Yuen, “Graph-attention-based reinforcement learning for trajectory design and resource assignment in multi-UA V-assisted communication,”IEEE Internet of Things Journal, vol. 11, no. 16, pp. 27 421–27 434, 2024

    2024

  11. [19]

    Joint power and 3D trajectory optimization for UA V-enabled wireless powered communi- cation networks with obstacles,

    H. Pan, Y . Liu, G. Sun, J. Fan, S. Liang, and C. Yuen, “Joint power and 3D trajectory optimization for UA V-enabled wireless powered communi- cation networks with obstacles,”IEEE Transactions on Communications, vol. 71, no. 4, pp. 2364–2380, 2023

    2023

  12. [20]

    Hy- bridRAG: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction,

    B. Sarmah, D. Mehta, B. Hall, R. Rao, S. Patel, and S. Pasquali, “Hy- bridRAG: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction,” inProceedings of the 5th ACM International Conference on AI in Finance, 2024, p. 608–616

    2024

  13. [21]

    HybGRAG: Hybrid retrieval- augmented generation on textual and relational knowledge bases,

    M.-C. Lee, Q. Zhu, C. Mavromatis, Z. Han, S. Adeshina, V . N. Ioannidis, H. Rangwala, and C. Faloutsos, “HybGRAG: Hybrid retrieval- augmented generation on textual and relational knowledge bases,”arXiv preprint arXiv:2412.16311, 2024

    2024 arXiv

  14. [22]

    Large language models and artificial intelligence generated content technologies meet communication networks,

    J. Guo, M. Wang, H. Yin, B. Song, Y . Chi, F. R. Yu, and C. Yuen, “Large language models and artificial intelligence generated content technologies meet communication networks,”IEEE Internet of Things Journal, vol. 12, no. 2, pp. 1529–1553, 2025

    2025

  15. [23]

    Diffusion policies as an expressive policy class for offline reinforcement learning,

    Z. Wang, J. J. Hunt, and M. Zhou, “Diffusion policies as an expressive policy class for offline reinforcement learning,” inInternational Confer- ence on Learning Representations, 2023

    2023

  16. [24]

    Efficient diffusion policies for offline reinforcement learning,

    B. Kang, X. Ma, C. Du, T. Pang, and S. Yan, “Efficient diffusion policies for offline reinforcement learning,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 67 195–67 212

    2023

  17. [25]

    Diffusion-based dynamic contract for federated AI agent construction in mobile metaverses,

    J. Wen, J. Kang, Y . Zhang, Y . Zhong, D. Niyato, J. Xu, J. Tang, and C. Yuen, “Diffusion-based dynamic contract for federated AI agent construction in mobile metaverses,”arXiv preprint arXiv:2504.14326, 2025

    2025

  18. [26]

    Generative diffusion-based contract design for efficient AI twin migration in vehicular embodied AI networks,

    Y . Zhong, J. Kang, J. Wen, D. Ye, J. Nie, D. Niyato, X. Gao, and S. Xie, “Generative diffusion-based contract design for efficient AI twin migration in vehicular embodied AI networks,”IEEE Transactions on Mobile Computing, vol. 24, no. 5, pp. 4573–4588, 2025

    2025

  19. [27]

    Optimizing resource allocation for multi-modal semantic communication in mobile AIGC networks: A diffusion-based game approach,

    J. Liu, M. Xiao, J. Wen, J. Kang, R. Zhang, T. Zhang, D. Niyato, W. Zhang, and Y . Liu, “Optimizing resource allocation for multi-modal semantic communication in mobile AIGC networks: A diffusion-based game approach,”IEEE Transactions on Cognitive Communications and Networking...

    2025

  20. [28]

    AoI-aware scheduling for air-ground collaborative mobile edge computing,

    Z. Qin, Z. Wei, Y . Qu, F. Zhou, H. Wang, D. W. K. Ng, and C.-B. Chae, “AoI-aware scheduling for air-ground collaborative mobile edge computing,”IEEE Transactions on Wireless Communications, vol. 22, no. 5, pp. 2989–3005, 2023

    2023

  21. [29]

    Multi-agent deep reinforcement learning for task offloading in UA V-assisted mobile edge computing,

    N. Zhao, Z. Ye, Y . Pei, Y .-C. Liang, and D. Niyato, “Multi-agent deep reinforcement learning for task offloading in UA V-assisted mobile edge computing,”IEEE Transactions on Wireless Communications, vol. 21, no. 9, pp. 6949–6960, 2022

    2022

  22. [30]

    Delivery by drone: An evaluation of un- manned aerial vehicle technology in reducing CO2 emissions in the delivery service industry,

    A. Goodchild and J. Toy, “Delivery by drone: An evaluation of un- manned aerial vehicle technology in reducing CO2 emissions in the delivery service industry,”Transportation Research Part D: Transport and Environment, vol. 61, pp. 58–67, 2018

    2018

  23. [31]

    Retrieval-augmented generation for mobile edge computing via large language model,

    R. Ren, Y . Wu, X. Zhang, J. Ren, Y . Shen, S. Wang, and K.-F. Tsang, “Retrieval-augmented generation for mobile edge computing via large language model,”arXiv preprint arXiv:2412.20820, 2024

    2024 arXiv

  24. [32]

    Hybrid RAG-empowered multimodal LLM for secure data management in Internet of Medical Things: A diffusion- based contract approach,

    C. Su, J. Wen, J. Kang, Y . Wang, Y . Su, H. Pan, Z. Zhong, and M. Shamim Hossain, “Hybrid RAG-empowered multimodal LLM for secure data management in Internet of Medical Things: A diffusion- based contract approach,”IEEE Internet of Things Journal, vol. 12, no. 10, pp. 13 428–...

    2025

  25. [33]

    Diffusion-model-based incentive mechanism with prospect theory for edge AIGC services in 6G IoT,

    J. Wen, J. Nie, Y . Zhong, C. Yi, X. Li, J. Jin, Y . Zhang, and D. Niyato, “Diffusion-model-based incentive mechanism with prospect theory for edge AIGC services in 6G IoT,”IEEE Internet of Things Journal, vol. 11, no. 21, pp. 34 187–34 201, 2024

    2024

  26. [34]

    Multi- agent DRL for multi-objective twin migration routing with workload prediction in 6G-enabled IoV,

    P. Yin, W. Liang, J. Wen, J. Kang, J. Chen, and D. Niyato, “Multi- agent DRL for multi-objective twin migration routing with workload prediction in 6G-enabled IoV,”arXiv preprint arXiv:2505.07290, 2025

    2025

  27. [35]

    Diffusion-based reinforcement learning via Q-weighted variational policy optimization,

    S. Ding, K. Hu, Z. Zhang, K. Ren, W. Zhang, J. Yu, J. Wang, and Y . Shi, “Diffusion-based reinforcement learning via Q-weighted variational policy optimization,”arXiv preprint arXiv:2405.16173, 2024

    2024 arXiv

  28. [36]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” inProceedings of the 35th International Conference on Machine Learning, vol. 80, 2018, pp. 1861–1870

    2018

  29. [37]

    Confidence-regulated generative diffusion models for reliable AI agent migration in vehicular metaverses,

    Y . Kang, J. Kang, J. Wen, T. Zhang, Z. Yang, D. Niyato, and Y . Zhang, “Confidence-regulated generative diffusion models for reliable AI agent migration in vehicular metaverses,”arXiv preprint arXiv:2505.12710, 2025

    2025 arXiv

  30. [38]

    Sustainable diffusion-based incentive mechanism for generative AI-driven digital twins in industrial cyber-physical systems,

    J. Wen, J. Kang, D. Niyato, Y . Zhang, and S. Mao, “Sustainable diffusion-based incentive mechanism for generative AI-driven digital twins in industrial cyber-physical systems,”IEEE Transactions on In- dustrial Cyber-Physical Systems, vol. 3, pp. 139–149, 2025

    2025

  31. [39]

    Collaboration in the sky: A distributed framework for task offloading and resource allocation in multi-access edge computing,

    Y . K. Tun, T. N. Dang, K. Kim, M. Alsenwi, W. Saad, and C. S. Hong, “Collaboration in the sky: A distributed framework for task offloading and resource allocation in multi-access edge computing,”IEEE Internet of Things Journal, vol. 9, no. 23, pp. 24 221–24 235, 2022

    2022

  32. [40]

    Task offloading and trajectory optimization for secure communications in dynamic user multi-UA V MEC systems,

    Y . Zhang, Z. Kuang, Y . Feng, and F. Hou, “Task offloading and trajectory optimization for secure communications in dynamic user multi-UA V MEC systems,”IEEE Transactions on Mobile Computing, vol. 23, no. 12, pp. 14 427–14 440, 2024

    2024

  33. [41]

    RAGChecker: A fine-grained framework for diagnosing retrieval-augmented generation,

    D. Ru, L. Qiu, X. Hu, T. Zhang, P. Shi, S. Chang, C. Jiayang, C. Wang, S. Sun, H. Li, Z. Zhang, B. Wang, J. Jiang, T. He, Z. Wang, P. Liu, Y . Zhang, and Z. Zhang, “RAGChecker: A fine-grained framework for diagnosing retrieval-augmented generation,” inAdvances in Neural Inform...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.