Pith. sign in

REVIEW 3 major objections 7 minor 42 references

A base station that receives sparse channel and content-preference feedback can prioritize important image regions and lower semantic reconstruction error via collaborative multi-agent learning.

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 · grok-4.5

2026-07-31 03:38 UTC pith:3KSEWJC2

load-bearing objection Solid systems/MARL paper on joint content-aware SC scheduling plus sparse dual feedback; headline % gains are on a shaped reward, not raw weighted MSE, so treat the abstract numbers carefully. the 3 major comments →

arxiv 2607.25011 v1 pith:3KSEWJC2 submitted 2026-07-27 cs.NI

Optimization of Collaborative Semantic Communication Network Performance with Channel and Content Preference Feedback

Defeng Zhou , Dongyu Wei , Siyao Li , Mingzhe Chen This is my paper
classification cs.NI
keywords semantic communicationimage transmissioncontent preference feedbackchannel state informationmulti-agent reinforcement learningactor-criticvalue decompositiondynamic neighborhood construction
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.

Semantic image links today usually treat every region as equally important, so bandwidth is spent on sky and background while cars, signs, or faces suffer. This paper builds a system in which a base station splits each image into content-based sub-images, encodes them, and ships them over wireless sub-channels, while users must decide—under power and bandwidth limits—whether to spend each uplink slot on channel-state feedback, on a short list of preferred content categories, or on silence. The goal is to minimize preference-weighted mean-square error between the original and the regenerated image. The authors solve the joint sub-channel, power, and feedback-selection problem with VDAC-DNC: an actor-critic scheme that approximates large discrete actions by a continuous policy, then searches only a small local neighborhood for the discrete choice with highest Q-value. Simulations show several-percent gains over standard multi-agent actor-critic, larger gains when feedback is removed, and multi-fold faster training from the neighborhood search.

Core claim

When users cooperatively choose when to send limited CSI versus content-preference feedback, and the base station jointly allocates sub-channels and power, semantic-weighted reconstruction error falls relative to equal-importance transmission and to non-cooperative learners; VDAC-DNC recovers these policies efficiently by continuous action approximation plus local Q-value neighborhood search rather than full discrete enumeration.

What carries the argument

VDAC-DNC (value-decomposition actor-critic with dynamic neighborhood construction): value decomposition aggregates local Q-functions into a team reward, the actor emits a continuous action distribution that is rounded to a discrete seed, and DNC expands only a small neighbor set around that seed to pick the action with maximum Q-value, avoiding traversal of the full discrete allocation space.

Load-bearing premise

Users obtain perfect channel state from pilots and can therefore send accurate CSI feedback and compute exact uplink rates; if that CSI is noisy or stale, both the feedback content and the training reward become wrong.

What would settle it

Repeat the multi-user image trials with noisy or delayed pilot-based CSI at the users and check whether the reported reward and weighted-MSE gains over multi-agent QAC and no-feedback baselines shrink or reverse.

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

If this is right

  • Sparse joint CSI and preference feedback measurably lowers semantic-weighted MSE versus CSI-only, preference-only, or silent operation.
  • Collaborative feedback timing keeps uplink load near a stable few messages per slot, avoiding the congestion or under-use that independent learners produce.
  • Dynamic neighborhood construction yields multi-fold training speedup while preserving final reward as the number of users grows.
  • Even with heterogeneous user power budgets, coordination still improves team reward because high-power users can report more often.
  • Sub-image scheduling that reacts to received preference weights can protect important objects on better channels while deprioritizing background.

Where Pith is reading between the lines

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

  • The continuous-policy-plus-local-neighborhood trick is reusable for other large combinatorial wireless decisions such as joint association and beam selection.
  • Top-D category preference vectors are a natural bridge to existing ROI or attention codecs; fusing them could shrink the uplink feedback payload further.
  • Deployment will need an explicit model of imperfect uplink CSI inside the state, because the perfect-pilot assumption is the first thing real channels violate.

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

3 major / 7 minor

Summary. The paper proposes a semantic communication framework in which a BS transmits content-segmented sub-images to multiple users, and users cooperatively decide when to send CSI feedback, content-preference feedback, or remain silent, under uplink bandwidth and power budgets. The authors formulate a semantic-weighted MSE minimization problem over sub-channel allocation, feedback selection, and user power (Eq. 6), and solve it with a value-decomposition actor-critic method augmented by a "dynamic neighborhood construction" (DNC) module that rounds the continuous actor output and runs a simulated-annealing search over a small discrete neighborhood using the critic's Q-values. A convergence lemma (Lemma 1) is given, and simulations compare against MAQAC, independent Q, an ablation without DNC, and one-shot transmission, reporting reward gains of up to 5.04%/18.55% and a 2.9x–5.3x training speedup from DNC.

Significance. If the results hold, the paper makes a useful contribution: the explicit tradeoff between spending scarce uplink resources on CSI versus content-preference feedback is, to my knowledge, not treated in prior semantic-communication work ([13]–[17] consider one or neither), and the problem is genuinely multi-agent because simultaneous feedback transmissions share the uplink band. The DNC mechanism — using a continuous policy to seed a small discrete Q-maximizing neighborhood search — is a reasonable and clearly described approach to large discrete action spaces, and the paper ships a non-trivial set of ablations (feedback type in Fig. 8, DNC search range in Fig. 9, device heterogeneity in Fig. 10, convergence-time measurements in Fig. 5). However, the headline quantitative claims are stated as percentages of a composite RL reward rather than of the paper's own objective (semantic-weighted MSE, Eq. 6), which currently limits how much weight the empirical claims can bear.

major comments (3)
  1. [§V, Figs. 3–4, 8–10; Eq. (9); Table I] The headline claims (abstract and §V: 'improve performance by up to 5.04% and 18.55%') are percentages of the accumulated composite reward r = (D^B − D^A) − κ_R e_R − κ_F e_F (Eq. 9), not of the stated objective in Eq. (6). The reward mixes an MSE-difference term with integer counts of delay violations, so these percentages depend on the relative scale set by κ_R and κ_F — yet κ_R and κ_F appear nowhere in Table I, are never tuned, and are never ablated. Moreover, a percentage of an accumulated reward stream is not well-defined if the accumulated reward is near zero or changes sign, and the y-axes of Figs. 3–4 do not let the reader rule this out. The only raw-MSE comparison in the paper (Table II) is against one-shot transmission, a much weaker baseline, where gains are 55–84% — which sits oddly next to single-digit reward gains against the real baselines. The authors should (i) report t
  2. [§IV.A, Lemma 1 and its proof; Eqs. (22)–(29); Algorithm 1] The convergence argument has two gaps. (1) Condition (i), E[Q_G(o,a)] = Q^{πθ}(o,a), is an exact-unbiasedness assumption on the function approximator; the proof discharges it with 'the error introduced by the Q function approximation decreases as the neural networks are trained,' which restates the assumption rather than establishing it, and the cited result [34] (COMA) does not assume exact unbiasedness. (2) More seriously, the reduction to a single-agent policy gradient in Eqs. (27)–(29) is valid only if the executed joint action is drawn from the factored policy πθ. It is not: the BS's executed action is the output of the DNC simulated-annealing search (Algorithm 1), i.e., a Q-maximizing neighbor of the actor's sample, not the sample itself. The executed policy is therefore a critic-dependent distortion of πθ, and the identity ∇θJ = E[∇θ log πθ · Q^{πθ}] does not apply to it without a
  3. [§II.B.2, Eq. (3); §IV.B] The uplink model assumes users obtain perfect CSI via pilot-assisted estimation, while the downlink rate without fresh CSI collapses to a fixed outage-capacity formula (Eq. 2). This asymmetry drives the entire feedback tradeoff: the value of a CSI-feedback action is computed against an exact uplink rate, and the reward that trains VDAC-DNC is evaluated under the same assumption. Since §IV.B claims the method 'can be deployed in practical scenarios,' the paper should include at least one robustness experiment or analysis with noisy or stale uplink CSI (e.g., estimation error variance swept over a plausible range), showing whether the learned feedback policy and the reported gains survive. The internal comparisons against baselines are unaffected (all share the channel model), so this is a correctness-risk concern about external validity rather than an internal inconsistency — but it curre
minor comments (7)
  1. [Table I] Several quantities needed to reproduce the simulations are missing: κ_R, κ_F, the discount factor γ, learning rates λ_a, λ_c, λ_h, slot durations Δt and Δt_f, K and T, and the DNC parameters y, c_β, c_k (l and ε are given). N_0 is listed twice with different formatting. The image dataset used for the sub-image experiments is never named anywhere in §V — please state it.
  2. [§II.B.1, Eq. (2)] The 'otherwise' branch of Eq. (2) is described as the standard outage-capacity formulation, but the expression log2(1 − P^D d_v^{-2} ln(1−ξ)/(I+N_0)) does not match the standard Rayleigh outage rate log2(1 + F^{-1}(ξ)·SNR) in [29]; it can also go negative inside the log for plausible parameter ranges. Please verify the formula or the derivation.
  3. [§III.B.2, Eqs. (16)–(18)] Eq. (16) defines L_c as the expected TD error (unsquared), and Eqs. (17)–(18) then take gradients of (1/2)L_c^2 — i.e., the square of an expectation rather than the expected squared TD error. Presumably the intended loss is E[(TD error)^2]; please correct the notation. Similarly, Eqs. (19)–(20) multiply the score function by L_c without explaining that the TD error serves as the advantage estimate; a sentence would help.
  4. [§V, Fig. 10 discussion] The text says that as the fraction of low-power users increases from 20% to 80%, the proposed method's reward increases by up to 11.84%, and explains this via high-power devices transmitting feedback more frequently. As written this is internally contradictory (or the x-axis/percentages are mislabeled). Please clarify.
  5. [§III.A, BS Actions] δ_{n,v}(t) ∈ {0, J_{k,v}} should presumably be {0, 1, ..., J_{k,v}} (an index set), not the two-element set containing the count.
  6. [References] [24] and [40] are the same MADDPG paper cited twice. Several references are arXiv preprints ([15], [17], [21], [26], [42]); please update to published versions where available.
  7. [§V, Fig. 3 discussion] The 5.04%/18.55% figures are 'up to' values taken at episode 14000; given the shaded half-standard-deviation bands in Fig. 3, please report whether the gap is statistically significant at convergence (e.g., mean ± std of converged reward over seeds), rather than at a single episode count.

Circularity Check

0 steps flagged

No significant circularity: empirical MARL optimization paper whose claims are simulation comparisons against external baselines, not first-principles predictions forced by definition or self-citation.

full rationale

The paper formulates a concrete resource-allocation problem (Eq. 6: minimize semantic-weighted MSE subject to allocation, power, and delay constraints) and solves it with a hybrid actor-critic / value-decomposition algorithm plus a neighborhood search (DNC). The objective, channel model, reward shaping (Eq. 9), and algorithm components are independently specified; none is defined in terms of the reported percentage gains. Comparative claims (5.04%, 18.55%, etc.) are empirical outcomes of Monte-Carlo training against external baselines (MAQAC / MADDPG-style centralized critic, independent Q-learning, one-shot transmission), not quantities recovered by construction from fitted parameters. Self-citations (e.g., prior Chen/Zhou/Wei MARL or semantic-comm works) appear only as related-work context or standard policy-gradient rewrite steps; none supplies a uniqueness theorem or ansatz that forces the present result. Hyperparameters (κ_R, κ_F, DNC cooling schedule, l) are author-chosen free parameters that affect the optimized policy, but they are not fitted so that any equation equals the claimed improvement. Metric mismatch between composite reward and raw weighted MSE is a reporting/validity concern, not circularity. Derivation chain is self-contained engineering + simulation; score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central performance claims rest on standard wireless fading/outage models, a perfect-uplink-CSI assumption, a particular semantic segmentation+VAE+compression stack treated as given, cooperative MARL credit assignment via value decomposition, and several hand-set penalty and DNC search hyperparameters that shape the learned policy. No new physical entity is postulated; the invented pieces are algorithmic (VDAC-DNC, DNC neighborhood) and the semantic-weighted MSE objective with dual feedback types.

free parameters (5)
  • κR, κF (delay penalty weights in reward)
    Hand-chosen multipliers on sub-image and feedback delay violations in (9); they steer the policy away from pure MSE minimization and directly affect reported rewards.
  • DNC search hyperparameters (l, y, ε, cβ, ck) = l=10 in Table I; ablations use l∈{0,1,2,4}
    Neighborhood range l and SA cooling schedule control which discrete actions are evaluated; Fig. 9 shows reward depends on l (best near l=2).
  • Actor/critic/MLP learning rates λa, λc, λh and network sizes
    Required for Lemma 1’s Robbins–Monro story and for any reproduced training curve; not fully specified numerically beyond symbolic rates.
  • Content preference top-D and importance weights ω = D=4
    D=4 in Table I; how ω are sampled or labeled per user/image is not fully specified but enters the objective linearly.
  • Simulation scenario knobs (V, N, pmax, ξ, interference range, distances) = V=15, N=6, pmax=5W, ξ=0.2, etc.
    Define the environment in which the 5.04%/18.55% gaps are measured; Table I and §V.
axioms (5)
  • domain assumption Users obtain perfect instantaneous CSI via pilots and can form exact CSI feedback vectors cv(t).
    Stated in II.B.2; underpins both uplink rate (3) and the information content of CSI feedback used by the BS state.
  • domain assumption Downlink without updated CSI follows the standard outage-capacity log expression with fixed outage probability ξ (Eq. 2).
    II.B.1; determines rate and thus delay constraint (6g) whenever feedback is missing or late.
  • domain assumption Sub-image extraction (InternImage-style), VAE encoding, and compression/decompression are fixed and sufficiently good that residual error is dominated by scheduling and packet loss.
    II.A treats the codec stack as given; optimization only chooses allocation/feedback/power.
  • standard math Global Q is well approximated by sum or QMIX-style monotonic mixing of local Qs, and multi-agent policy gradient reduces to single-agent form under factorization (Lemma 1).
    III–IV; standard MARL assumptions; DNC asserted not to break convergence because it only selects within the original discrete set.
  • ad hoc to paper One sub-image occupies at most one sub-channel per slot; each user sends at most one feedback type per slot; total uplink band is split equally among simultaneous feedback users g(t).
    Constraints (6c)–(6e) and rate (3); modeling choices that define the cooperation problem the algorithm solves.
invented entities (2)
  • VDAC-DNC algorithm (value-decomposition actor-critic + dynamic neighborhood construction) no independent evidence
    purpose: Scale multi-agent learning to large discrete BS allocation spaces while preserving cooperative credit assignment for feedback and power.
    Core methodological contribution; continuous actor + critic-guided SA-like local search + VDN/QMIX mixer. No claim of a new physical object.
  • Semantic-weighted MSE objective with joint CSI/content feedback selection no independent evidence
    purpose: Scalarize image importance and radio reliability into one team reward under delay and power caps.
    Problem (6) and reward (9); standard MSE reweighted by preference, not a new physical quantity.

pith-pipeline@v1.2.0-grok45-kimik3 · 26442 in / 3936 out tokens · 73078 ms · 2026-07-31T03:38:19.448809+00:00 · methodology

0 comments
read the original abstract

Existing semantic communication frameworks treat and transmit all image regions with equal importance, which is not practical for real-world applications which may prioritize different content in an image. To address this issue, we propose a novel semantic communication framework that enables a transmitter to use limited channel and content feedback to prioritize the transmission of important image regions. In particular, in the proposed framework, a base station (BS) divides each image into sub-images, extracts their semantic information, and transmits them to users according to their preferences. The users will reconstruct the image based on the received sub-images and cooperatively decide when to send channel state information (CSI) or content-preference feedback under dynamic channels and limited resources. We formulate an optimization problem to minimize the semantic-weighted mean square error between the original image and the regenerated image by optimizing sub-channel allocation, users' power allocation, and feedback selection. To address this problem, a value decomposition actor- critic (AC) with dynamic neighborhood construction (VDAC-DNC) scheme is proposed. The proposed method combines AC with value decomposition networks to allow the BS to approximate discrete actions by a continuous action distribution, thus reducing the output dimension and improving training efficiency. The introduced DNC method further improves training efficiency by constructing a small discrete neighboring action space to search for an action with the maximum Q value, thus avoiding traversing the large discrete action space. Simulation results show that the proposed VDAC-DNC scheme can improve the performance by up to 5.04% and 18.55% compared to the standard multi-agent QAC method and the proposed method without feedback transmission.

Figures

Figures reproduced from arXiv: 2607.25011 by Defeng Zhou, Dongyu Wei, Mingzhe Chen, Siyao Li.

Figure 1
Figure 1. Figure 1: Channel and preference feedback enabled semantic communication [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of the proposed VDAC-DNC algorithm [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accumulated reward versus the number of users [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Convergence of the considered algorithm optimizes sub-image allocation and feedback selection using the designed RL method but does not use DNC to find better actions per iteration [32], 2) a DNC based MAQAC algorithm that consists of a centralized critic and decentralized actors [40], and 3) DNC-based independent Q algorithm in which each user utilizes a deep Q network (DQN) to optimize its transmit power… view at source ↗
Figure 6
Figure 6. Figure 6: Number of feedback transmitted over an episode [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of sub-images and feedback messages transmission in each step during an episode [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Convergence of the designed methods with different types of feedback [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The impact of device heterogeneity on the proposed framework [PITH_FULL_IMAGE:figures/full_fig_p012_10.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

42 extracted references · 4 linked inside Pith

  1. [1]

    Less data, more knowledge: Building next-generation semantic communication networks,

    C. Chaccour, W. Saad, M. Debbah, Z. Han, and H. Poor, “Less data, more knowledge: Building next-generation semantic communication networks,” IEEE Commun. Surv. Tutor ., vol. 27, no. 1, pp. 37–76, Feb. 2025

    2025

  2. [2]

    Cross-modal generative semantic communications for mobile aigc: Joint semantic encoding and prompt engineering,

    Y . Liu, H. Du, D. Niyato, J. Kang, Z. Xiong, S. Mao, P. Zhang, and X. Shen, “Cross-modal generative semantic communications for mobile aigc: Joint semantic encoding and prompt engineering,”IEEE Trans. Mob. Comput., vol. 23, no. 12, pp. 14871–14888, Aug. 2024

    2024

  3. [3]

    Task-oriented communications for nextg: End-to-end deep learning and AI security aspects,

    Y . Sagduyu, S. Ulukus, and A. Yener, “Task-oriented communications for nextg: End-to-end deep learning and AI security aspects,”IEEE Wireless Commun., vol. 30, no. 3, pp. 52–60, Jun. 2023

    2023

  4. [4]

    Task-oriented multi-user semantic communications for VQA,

    H. Xie, Z. Qin, and G. Li, “Task-oriented multi-user semantic communications for VQA,”IEEE Wireless Commun. Lett., vol. 11, no. 3, pp. 553–557, Mar. 2022

    2022

  5. [5]

    Semantic communication: A survey on research landscape, challenges, and future directions,

    T. M. Getu, G. Kaddoum, and M. Bennis, “Semantic communication: A survey on research landscape, challenges, and future directions,”Proc. IEEE, vol. 112, no. 11, pp. 1649–1685, Nov. 2024

    2024

  6. [6]

    Deep joint source-channel coding for wireless image transmission with adaptive rate control,

    M. Yang and H. Kim, “Deep joint source-channel coding for wireless image transmission with adaptive rate control,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Singapore, May 2022, pp. 5193–5197

    2022

  7. [7]

    Semantic communications for image recovery and classification via deep joint source and channel coding,

    Z. Lyu, G. Zhu, J. Xu, B. Ai, and S. Cui, “Semantic communications for image recovery and classification via deep joint source and channel coding,”IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 8388–8404, Jan. 2024

    2024

  8. [8]

    Energy efficient multi-modal probabilistic semantic communication (PSCom),

    J. Dai, J. Li, Z. Zhao, Z. Yang, Z. Zhang, and M. Shikh-Bahaei, “Energy efficient multi-modal probabilistic semantic communication (PSCom),” IEEE Trans. Green Commun. Netw., vol. 9, no. 4, pp. 1951–1963, Dec. 2025

    1951

  9. [9]

    Predictive and adaptive deep coding for wireless image transmission in semantic communication,

    W. Zhang, H. Zhang, H. Ma, H. Shao, N. Wang, and V . Leung, “Predictive and adaptive deep coding for wireless image transmission in semantic communication,”IEEE Trans. Wireless Commun., vol. 22, no. 8, pp. 5486–5501, Aug. 2023

    2023

  10. [10]

    WITT: A wireless image transmission transformer for semantic communications,

    K. Yang, S. Wang, J. Dai, K. Tan, K. Niu, and P. Zhang, “WITT: A wireless image transmission transformer for semantic communications,” inProc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), Rhodes Island, Greece, Jun. 2023, pp. 1–5

    2023

  11. [11]

    On the capacity region of the layered packet erasure broadcast channel with feedback,

    S. Li, D. Tuninetti, and N. Devroye, “On the capacity region of the layered packet erasure broadcast channel with feedback,” inProc. IEEE Int. Conf. Commun. (ICC), Shanghai, China, May 2019, pp. 1–6

    2019

  12. [12]

    On the stability region of the layered packet erasure broadcast channel with output feedback,

    S. Li, H. Seferoglu, D. Tuninetti, and N. Devroye, “On the stability region of the layered packet erasure broadcast channel with output feedback,” inProc. IEEE Inf. Theory Workshop (ITW), Visby, Sweden, Aug. 2019, pp. 1–5

    2019

  13. [13]

    SCAN: Semantic communication with adaptive channel feedback,

    G. Zhang, Q. Hu, Y . Cai, and G. Yu, “SCAN: Semantic communication with adaptive channel feedback,”IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 5, pp. 1759–1773, Oct. 2024

    2024

  14. [14]

    Transformer-aided wireless image transmission with channel feedback,

    H. Wu, Y . Shao, E. Ozfatura, K. Mikolajczyk, and D. Gündüz, “Transformer-aided wireless image transmission with channel feedback,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 11904–11919, Sep. 2024

    2024

  15. [15]

    Adaptive wireless image semantic transmission: Design, simulation, and prototype validation,

    J. Ding, P. Jiang, C. Wen, and S. Jin, “Adaptive wireless image semantic transmission: Design, simulation, and prototype validation,” 2024, [Online]. Available: https://arxiv.org/abs/2410.17536

    Pith/arXiv arXiv 2024

  16. [16]

    Semantic segmentation-based semantic communication system for image transmission,

    J. Wu, C. Wu, Y . Lin, T. Yoshinaga, L. Zhong, X. Chen, and Y . Ji, “Semantic segmentation-based semantic communication system for image transmission,”Digit. Commun. Netw., vol. 10, no. 3, pp. 519–527, Jun. 2024

    2024

  17. [17]

    Quantization- aware collaborative inference for large embodied ai models,

    Z. Lyu, M. Xiao, M. Skoglund, M. Debbah, and H. V . Poor, “Quantization- aware collaborative inference for large embodied ai models,” 2026, [Online]. Available: https://arxiv.org/abs/2602.13052

    arXiv 2026

  18. [18]

    Covert UA V data transmission via semantic communication: A DRL-driven joint position and power optimization method,

    R. Xu, G. Li, Z. Yang, J. Kang, X. Zhang, and J. Li, “Covert UA V data transmission via semantic communication: A DRL-driven joint position and power optimization method,” inProc. Int. Conf. Commun. China (ICCC), Hangzhou, China, Aug. 2024, pp. 66–71

    2024

  19. [19]

    Performance optimization for semantic communications: An attention- based reinforcement learning approach,

    Y . Wang, M. Chen, T. Luo, W. Saad, D. Niyato, H. V . Poor, and S. Cui, “Performance optimization for semantic communications: An attention- based reinforcement learning approach,”IEEE J. Select. Areas Commun., vol. 40, no. 9, pp. 2598–2613, Feb. 2022

    2022

  20. [20]

    Optimization of image transmission in cooperative semantic communication networks,

    W. Zhang, Y . Wang, M. Chen, T. Luo, and D. Niyato, “Optimization of image transmission in cooperative semantic communication networks,” IEEE Trans. Wireless Commun., vol. 23, no. 2, pp. 861–873, Feb. 2024

    2024

  21. [21]

    Semantic-aware resource management for C-V2X platooning via multi-agent reinforcement learning,

    W. Zhang, Q. Wu, P. Fan, K. Wang, N. Cheng, W. Chen, and K. B. Letaief, “Semantic-aware resource management for C-V2X platooning via multi-agent reinforcement learning,” 2025, [Online]. Available: https://arxiv.org/abs/2411.04672

    Pith/arXiv arXiv 2025

  22. [22]

    Challenges of real-world reinforcement learning: definitions, benchmarks and analysis,

    G. Dulac-Arnold et al., “Challenges of real-world reinforcement learning: definitions, benchmarks and analysis,”Mach. Learn., vol. 110, no. 9, pp. 2419–2468, Sep. 2021

    2021

  23. [23]

    Value-decomposition networks for cooperative multi- agent learning based on team reward,

    P. Sunehag et al., “Value-decomposition networks for cooperative multi- agent learning based on team reward,” inProc. Int. Conf. Auton. Agents Multiagent Syst., Stockholm, Sweden, Jul. 2018, pp. 2085–2087

    2018

  24. [24]

    Multi- agent actor-critic for mixed cooperative-competitive environments,

    R. Lowe, Y . Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch, “Multi- agent actor-critic for mixed cooperative-competitive environments,” in Proc. Neural Inf. Process. Syst. (NIPS), Long Beach, CA, USA, Dec. 2017, pp. 6382–6393

    2017

  25. [25]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), New Orleans, LA, USA, Jun. 2022, pp. 10684–10695

    2022

  26. [26]

    InternImage: Exploring large-scale vision foundation models with deformable convolutions,

    W. Wang et al., “InternImage: Exploring large-scale vision foundation models with deformable convolutions,” 2023, [Online]. Available: https://arxiv.org/abs/2211.05778. 14

    Pith/arXiv arXiv 2023

  27. [27]

    Scaling up your kernels to 31×31: Revisiting large kernel design in CNNs,

    X. Ding, X. Zhang, J. Han, and G. Ding, “Scaling up your kernels to 31×31: Revisiting large kernel design in CNNs,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), New Orleans, LA, USA, Jun. 2022, pp. 11953–11965

    2022

  28. [28]

    Swin Transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical vision transformer using shifted windows,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Montreal, QC, Canada, Oct. 2021, pp. 9992–10002

    2021

  29. [29]

    Tse and P

    D. Tse and P. Viswanath,Fundamentals of Wireless Communication, Cambridge University Press, 2005

    2005

  30. [30]

    Modeling and enhancement of the IEEE 802.11 RTS/CTS scheme in an error-prone channel,

    M. Yazid, D. Aissani, L. Bouallouche-Medjkoune, N. Amrouche, and B. Kamel, “Modeling and enhancement of the IEEE 802.11 RTS/CTS scheme in an error-prone channel,”F orm. Asp. Comput., Jan. 2015

    2015

  31. [31]

    A closer look at invalid action masking in policy gradient algorithms,

    S. Huang and S. Ontañón, “A closer look at invalid action masking in policy gradient algorithms,” inProc. Int. FLAIRS Conf., Jensen Beach, FL, USA, May 2022

    2022

  32. [32]

    Monotonic value function factorisation for deep multi- agent reinforcement learning,

    T. Rashid et al., “Monotonic value function factorisation for deep multi- agent reinforcement learning,”J. Mach. Learn. Res., vol. 21, no. 178, pp. 1–51, 2020

    2020

  33. [33]

    Kochenderfer and T

    M. Kochenderfer and T. Wheeler,Algorithms for Optimization, MIT Press, Cambridge, MA, USA, 2019

    2019

  34. [34]

    Counterfactual multi-agent policy gradients,

    J. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson, “Counterfactual multi-agent policy gradients,” inProc. AAAI Conf. Artif. Intell., New Orleans, Louisiana, USA, Apr. 2018, vol. 32

    2018

  35. [35]

    Policy gradient methods for reinforcement learning with function approximation,

    R. S Sutton, D. A McAllester, S. P Singh, and Y . Mansour, “Policy gradient methods for reinforcement learning with function approximation,” inProc. Adv. Neural Inf. Process. Syst., Cambridge, MA, USA, Nov. 1999, pp. 1057–1063

    1999

  36. [36]

    Learning adaptive jamming and beamforming for hybrid IRS-assisted secure NOMA transmissions,

    D. Zhou, L. Li, S. Gong, B. Gu, G. Chen, and D. Niyato, “Learning adaptive jamming and beamforming for hybrid IRS-assisted secure NOMA transmissions,”IEEE Trans. Commun., vol. 73, no. 11, pp. 12235–12247, Nov. 2025

    2025

  37. [37]

    Transformer- based collaborative reinforcement learning for fluid antenna system (FAS)-enabled 3D UA V positioning,

    X. Xu, H. Xu, D. Wei, W. Saad, M. Bennis, and M. Chen, “Transformer- based collaborative reinforcement learning for fluid antenna system (FAS)-enabled 3D UA V positioning,”IEEE J. Select. Areas Commun., vol. 44, pp. 1128–1143, 2026

    2026

  38. [38]

    Optimizing model splitting and device task assignment for deceptive signal-assisted private multi-hop split learning,

    D. Wei, X. Xu, Y . Liu, H. V . Poor, and M. Chen, “Optimizing model splitting and device task assignment for deceptive signal-assisted private multi-hop split learning,”IEEE J. Select. Areas Commun., vol. 44, pp. 1512–1528, 2026

    2026

  39. [39]

    Simulated annealing,

    D. Bertsimas and J. N. Tsitsiklis, “Simulated annealing,”Statist. Sci., vol. 8, no. 1, pp. 10–15, Feb. 1993

    1993

  40. [40]

    Multi- agent actor-critic for mixed cooperative-competitive environments,

    R. Lowe, Y . Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch, “Multi- agent actor-critic for mixed cooperative-competitive environments,” in Proc. Adv. Neural Inf. Process. Syst. (NIPS), Long Beach, CA, USA, Dec. 2017, vol. 30, pp. 6382–6393

    2017

  41. [41]

    Multi-agent reinforcement learning: Independent vs. cooperative agents,

    M. Tan, “Multi-agent reinforcement learning: Independent vs. cooperative agents,” inProc. Int. Conf. Mach. Learn. (ICML), Amherst, MA, USA, Jun. 1993, pp. 330–337

    1993

  42. [42]

    Resource allocation for semantic-aware net- works,

    L. Yan et al., “Resource allocation for semantic-aware net- works,”arXiv preprint arXiv:2201.06023, 2022, [Online]. Available: https://arxiv.org/abs/2201.06023

    Pith/arXiv arXiv 2022