Pith. sign in

REVIEW 5 major objections 5 minor 58 references

LLM-Empowered Agentic MAC Protocols: A Dynamic Stackelberg Game Approach

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that wireless MAC protocols can be made to emerge automatically by framing uplink scheduling as a dynamic multi-follower Stackelberg game, with a base-station leader and LLM-based user policies trained by PPO, and proves t

desk verdict A genuinely useful framing — LLM policies plus a dynamic multi-follower Stackelberg game for MAC emergence — but the headline numbers are unverifiable without code/error bars, and the convergence proof rests on an assumption that likely fails in the simulated environment. read the letter →

arxiv 2510.10895 v2 pith:DCGGBAYI submitted 2025-10-13 cs.AI

Renxuan Tan , Rongpeng Li , Fei Wang , Chenghui Peng , Shaoyun Wu , Zhifeng Zhao , Honggang Zhang This is my paper · ORCID
classification cs.AI MSC 91A6568T05
keywords MACprotocolsStackelberggamemulti-agentreinforcementlearningprotocolemergencelargelanguagemodelsPPOuplinkschedulingsemanticcommunication
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that wireless medium-access control (MAC) protocols can be made to emerge automatically by framing uplink scheduling as a dynamic multi-follower Stackelberg game: a base station leads, user devices follow, and both sides are LLM-based policies trained with PPO. The authors prove that an expected Stackelberg equilibrium exists and that the PPO update converges to it locally, under a weak-coupling assumption. In simulations, the learned protocol reports 77.6% higher throughput and 65.2% better fairness than conventional baselines, and transfers to networks with varying numbers of users without retraining. The significance is showing a path from static, hand-designed MAC to adaptive, semantic protocols.

What carries the argument

The central object is the dynamic multi-follower Stackelberg game (MFSG) instantiated with LLM policies. A base-station leader broadcasts a downlink control message (DCM) as its action; user followers respond with uplink control messages (UCMs) and transmission bitmaps, balancing their own transmission efficiency against consistency with the leader's intent. Protocol action grammar (PAG) constrains the LLM's output vocabulary to valid action tokens so generated protocols are executable. PPO with shared follower parameters and a leader/follower time-scale separation drives the learning; the proofs hinge on showing the follower subgame is a potential game and the update map is a contraction.

What would settle it

Compute the Jacobian of the follower best-response system (Eq. 29) at a converged policy in the simulated UDTS environment; if its off-diagonal blocks are not negligibly small relative to the diagonal Hessians, strict diagonal dominance fails and the contraction argument in Appendix C breaks. A cheaper indicator: run Algorithm 1 in a high-collision regime (e.g., 10 UEs, 3 RBGs) and check whether training diverges when the PPO update is applied without the weak-coupling-justified time-scale separation.

Watch

Extended reading notes

Core claim

The central claim is that protocol emergence for the uplink can be modeled as a multi-follower Stackelberg game whose players are LLM-based policies, and that this game-theoretic structure is what enables generalization: because the LLM processes variable-length message sequences, the same trained policy serves 3, 4, 5, and up to 10 users, while a symbolic PPO baseline degrades or needs retraining. The paper further claims that the follower subgame is an exact potential game, guaranteeing an equilibrium, and that the PPO gradient dynamics are a local contraction around a differential Stackelberg equilibrium, giving geometric convergence.

Load-bearing premise

The proofs assume that near the equilibrium the cross-agent second-order effects in any utility are negligible (Eq. 17); if that weak coupling fails, both the existence of the best-response function and the convergence contraction can break.

Editorial extensions

If this is right

  • A single trained LLM policy can orchestrate MAC for 2–10 users without architecture changes or retraining.
  • The learned protocol maintains near-optimal Jain's fairness index (reported above 0.9 across scenarios).
  • The framework reports consistent throughput gains over retrained and non-retrained MARL baselines and an adaptive heuristic, with largest gains under high contention.
  • The existence of an expected Stackelberg equilibrium means the protocol design problem is at least well-posed under the stated assumptions.
  • Changing the number of resource block groups (3→8) does not break the policy, unlike the symbolic baseline whose throughput paradoxically falls.

Reading between the lines

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

  • A testable extension beyond this paper: in a strongly interfering regime, compute the off-diagonal Hessian blocks during training; if they are non-negligible, the contraction proof would require a modified time-scale separation.
  • If the weak-coupling assumption holds more broadly, the same leader-follower structure could be transplanted to other hierarchical wireless tasks such as slicing, power control, or multi-cell scheduling.
  • The use of natural-language control messages opens the door to human-in-the-loop auditing: operators could read UCMs and DCMs to understand why the BS and UEs made their choices.
  • The complexity analysis implies that inference cost grows with the number of users (prompt length), so real deployments might trade the reported throughput gains against added latency at the edge.
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

5 major / 5 minor

Summary. The paper proposes an LLM-empowered multi-agent reinforcement learning framework for uplink MAC protocol emergence. The base station and user equipments are modeled as a leader and followers in a dynamic multi-follower Stackelberg game, with Llama-style LLMs as policy networks constrained by a Protocol Action Grammar and trained with PPO. The authors claim to prove existence of an expected Stackelberg equilibrium and local geometric convergence of the learning dynamics, and report simulation results with 77.6% throughput and 65.2% fairness improvements over baselines, plus generalization to 2–10 users without retraining.

Significance. If fully established, the paper would be a timely contribution: the combination of a hierarchical Stackelberg game with LLM policies for protocol emergence is novel, and the variable-length prompt formulation is a natural way to handle dynamic user counts. PAG is a sensible engineering mechanism for reliable action generation, and the complexity analysis is a useful addition. The theoretical guarantees, however, are a central part of the paper's stated contribution and are not currently supported: the proofs rest on an unverified and structurally questionable weak-coupling assumption, analyze a gradient system that differs from the clipped PPO algorithm actually executed, and contain gaps in the existence argument. The empirical results are plausible but not independently verifiable from the manuscript. The core idea has merit, but the claimed guarantees and empirical evidence require substantial repair.

major comments (5)
  1. [§IV.E, Eq. (17); Apps. A and C] Assumption 2 is the keystone of both theorems. It is used in Appendix A to make the follower FOC Jacobian J_G strictly diagonally dominant and in Appendix C to make M(Θ*) negative definite. No numerical or analytical evidence is provided for Eq. (17) in the simulated OFDMA environment. In fact, the follower payoff in Eq. (24) is a product over other UEs' selection indicators, so the cross second derivative ∂²E[J_i]/∂θ_i∂θ_j is proportional to another UE's policy derivative and is not negligible in high-contention regimes. If Eq. (17) fails, the existence and convergence proofs collapse. I request a direct verification of Eq. (17) on the actual task (e.g., Hessian norms along training trajectories) or a replacement of this assumption with a condition that is actually satisfied by the problem.
  2. [§IV.D–E, Eq. (10) vs. Eq. (16)] Theorem 2 concerns the dynamical system Θ^{e+1}=Θ^e+α_bΩ(Θ^e) with Ω=∇ΘJ. Algorithm 1, however, updates actors with the clipped PPO surrogate in Eq. (10), which includes clipping, a KL penalty, an entropy bonus, GAE advantage estimates, a critic, and Adam with mini-batches. The clipped objective is not an exact gradient of the expected utilities J_i,u or J_b. Thus the convergence guarantee in Theorem 2 does not apply to the algorithm actually executed. The authors need to either prove that the PPO update acts as gradient ascent on the game utilities (with explicit treatment of the clipping/penalty terms) or state and prove a convergence result for the actual update rule.
  3. [Appendix A, Eqs. (27)–(29); Definition 2] There are two gaps in the existence proof. First, Eq. (28) defines G using the stage utility F_{i,t,u}, whereas the DSE condition in Definition 2 is ∇_{θ_{i,u}}J_{i,u}=0 for the expected discounted utility. The implicit function theorem is therefore applied to the wrong function. Second, the exact potential game argument only shows that the follower best-response correspondence is nonempty; it does not imply that h(θ_b) is single-valued and C¹. The IFT gives only a local branch around a known root. The continuity of the leader's objective and the Weierstrass step of Theorem 1 are not established as written.
  4. [Corollary 1 and App. D, Eqs. (34)–(36)] Corollary 1 claims almost sure convergence for a stochastically varying number of followers, but the proof analyzes deterministic averaged dynamics in Eq. (35) and shows contraction of the expected map. Algorithm 1 performs stochastic gradient updates on finite mini-batches with random UE sets; almost sure convergence requires a stochastic-approximation argument (step-size conditions, martingale noise control), which is absent. The contraction of ¯M(Θ*) does not, by itself, imply almost sure convergence of the stochastic iterates.
  5. [§V, Figs. 4–9; Table III] The central empirical claim is not independently verifiable from the manuscript. The text states that results are averaged over five independent runs, but no error bars, confidence intervals, seeds, or statistical tests are reported, and no code or data are released. The headline 77.6%/65.2% figures are point estimates from one configuration. Additionally, Table III lists e_max=2000 while Section V.A states 3000 epochs. Please include variance information, random seeds, and either code or detailed per-run logs.
minor comments (5)
  1. [Appendix A, Eq. (25)] The potential function E is said to equal network throughput, but the formula counts an RBG as productive whenever at least one UE selects it, regardless of collisions. The exact-potential property may still hold, but the physical interpretation should be corrected or clarified.
  2. [Figs. 4–6] Several figures appear to have rendering/font problems (e.g., missing axis labels, garbled tick labels). Please provide vector graphics with readable axis names and units.
  3. [Appendix A, Lemma 1] Reference [47] is cited for the Weierstrass theorem, but the cited paper is about the Stone–Weierstrass theorem and neural networks. Use a standard real-analysis reference for the extreme value theorem.
  4. [§IV.C, Fig. 3] The PAG vocabulary W_b^- is written as {"0","1",...,"I_t"}; the token "I_t" is unclear. Please clarify whether it denotes an index token or a special separator.
  5. [§V.A, Table III] The table does not list an entry for T (episode length), although T=24 is used in the text and Eq. (21). Also reconcile the e_max inconsistency noted above.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical claims are conditioned on explicitly stated assumptions and the empirical results are independently simulated.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The game-theoretic formulation (Eqs. 2–6), the PPO update rule (Eq. 16), and the convergence analysis (Theorems 1–2, Appendices A–D) do not redefine any input as a prediction. Theorem 1 uses a potential-game argument plus the implicit function theorem; Theorem 2 uses a contraction-mapping argument with a Schur complement. The weak-coupling Assumption 2 (Eq. 17) is an explicit sufficient condition, not a conclusion smuggled in as an input. The choice of learning rate α_b = κ1/κ2 and the selection of an 'appropriate' ι_u in Appendix C are constructive choices that establish existence of parameter values under which the contraction holds; this is a standard style of local convergence theorem, not a fitted parameter being relabeled as a prediction. The self-citation [10] is only contextual and does not carry a load-bearing argument. The simulation results are generated by an external environment and are not used to fit or rename the theoretical quantities. The only concern—whether Assumption 2 actually holds in the simulated OFDMA MAC environment—is an empirical/correctness risk, not a circularity here. No circular step can be exhibited, so the score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical or mathematical entities; it composes known components: LLM policies, PPO, constrained decoding (PAG), and a Stackelberg game. The main free choices are hand-set utility weights, the theoretical time-scale separation and learning-rate selection, and the undisclosed meta-prompts. The load-bearing axioms are Assumption 2 and the implicit equivalence between PPO updates and ideal policy-gradient ascent.

free parameters (6)
  • ϱ1, ϱ2 (follower utility weights) = 8, 10
    Hand-set in Eq. (2)/Table III; determine the UE's trade-off between successful transmission and compliance with the BS's DCM, directly shaping the emergent protocol.
  • ε (leader fairness weight) = 5
    Hand-set in Eq. (5)/Table III; controls the leader's trade-off between average throughput and Jain fairness index.
  • ι_u (follower time-scale separation) = not specified
    Introduced in Definition 3; Appendix C requires an 'appropriate choice of ι_u' to make the Schur complement negative definite, i.e., to force the convergence proof.
  • α_b (leader learning rate, set to κ1/κ2) = defined via singular values of M(Θ*) and J_Ω
    Theorem 2 requires this specific rate to guarantee contraction; it depends on unknown equilibrium quantities, so it functions as a tuning parameter rather than a constructively computed value.
  • Actor temperature schedule = linear decay 3.0 -> 0.3
    Table III; hand-chosen exploration schedule that affects policy entropy during training.
  • Meta-prompts ψ_b, ψ_u = not provided
    Section IV-B; hand-crafted natural-language prompts defining agent roles and task. Content is not disclosed, but they are a likely major determinant of LLM behavior.
assumptions (5)
  • standard math Assumption 1: Θ_b and Θ_u are non-empty compact sets and utilities are C^2-smooth.
    Section IV.E; standard regularity condition for gradient-based learning, typical in the cited learning-in-games literature.
  • ad hoc to paper Assumption 2: weak coupling near equilibrium, i.e., cross second-order derivatives between distinct agents are negligible vs. own-policy Hessians (Eq. 17).
    Section IV.E; asserted as 'a reasonable simplification' and used in both Theorem 1 (via diagonal dominance) and Theorem 2 (via Lemma 4). No verification from the system model is provided.
  • domain assumption The follower subgame is an exact potential game for each stage, and this extends to the dynamic stochastic game.
    Appendix A proves a potential function for the single-slot interactive payoff (Eq. 24-26), but the follower utilities J_i,u are discounted sums over T slots with state transitions; the existence of an NE in the full dynamic subgame is assumed rather than derived.
  • ad hoc to paper PPO's clipped surrogate objective (Eq. 10) is equivalent to gradient ascent on the agents' expected utilities (Definition 3, Eq. 16).
    Section IV.E; the convergence proof models the update as Θ ← Θ + α∇J, ignoring clipping, KL penalty, entropy bonus, and the learned critic. No argument is given that PPO's update direction equals the assumed gradient field.
  • ad hoc to paper The leader's best-response function h(θ_b) is single-valued and continuous via the implicit function theorem, given strict diagonal dominance of J_G.
    Appendix A; diagonal dominance is asserted 'by Assumption 2 and the follower's SOCs', making the existence of a continuous best response contingent on an unverified condition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Empowered Agentic MAC Protocols: A Dynamic Stackelberg Game Approach." pith.science (2026). https://pith.science/paper/DCGGBAYI

@misc{pith2026251010895,
  author       = {Pith},
  title        = {Pith review of: LLM-Empowered Agentic MAC Protocols: A Dynamic Stackelberg Game Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DCGGBAYI}},
  note         = {Machine review of arXiv:2510.10895}
}
read the original abstract

Medium Access Control (MAC) protocols, essential for wireless networks, are typically manually configured. While deep reinforcement learning (DRL)-based protocols enhance task-specified network performance, they suffer from poor generalizability and resilience, demanding costly retraining to adapt to dynamic environments. To overcome this limitation, we introduce a game-theoretic LLM-empowered multi-agent DRL (MARL) framework, in which the uplink transmission between a base station and a varying number of user equipments is modeled as a dynamic multi-follower Stackelberg game (MFSG), capturing the network's natural hierarchical structure. Within this game, LLM-driven agents, coordinated through proximal policy optimization (PPO), synthesize adaptive, semantic MAC protocols in response to network dynamics. Protocol action grammar (PAG) is employed to ensure the reliability and efficiency of this process. Under this system, we further analyze the existence and convergence behavior in terms of a Stackelberg equilibrium by studying the learning dynamics of LLM-empowered unified policies in response to changing followers. Simulations corroborate that our framework achieves a 77.6% greater throughput and a 65.2% fairness improvement over conventional baselines. Besides, our framework generalizes excellently to a fluctuating number of users without requiring retraining or architectural changes.

Figures

Figures reproduced from arXiv: 2510.10895 by the authors.

Figure 1
Figure 1. Scenario of interest: A BS serves dynamic UEs within an OFDMA [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The workflow of our LLM-empowered MFSG within a TTI. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. LLM policy action generation with PAG. modality. As depicted in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Performance comparison of throughput and fairness index under various packet arrival rates and network sizes, where [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 7
Figure 7. Figure 7: Performance comparison of throughput and fairness index with a [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Performance scalability with an increasing number of UEs. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Impact of TBLER on system throughput and fairness index. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Proof roadmap for existence theorem of other follower actions. To construct the potential function, we then reformulate this interactive utility F ′ i,t,u as F ′ i,t,u(a bit; at,b, I) = X M m=1 a bit i,t,u,m Y j∈It,j̸=i (1−a bit j,t,u,m)Nm  , (24) where a bit = (a b…
Figure 11
Figure 11. Figure 11: Proof roadmap for convergence theorem. B. Existence of an Optimal Universal Leader Policy We next analyze the leader’s problem as shown in Eq. (6), which can be reformulated as the parameterized value function form max θb∈Θb Jb(θb) = EI∼D(I) [Vb(θb, h(θb), I)] , (27) …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 11 linked inside Pith

  1. [1]

    Resilient-native and intelligent nextG systems,

    M. Bennis, “Resilient-native and intelligent nextG systems,” 2025. [Online]. Available: https://arxiv.org/abs/2506.12795 12

    arXiv 2025

  2. [2]

    Foundation model enhanced mul- tiple access in heterogeneous networks,

    M. Han, X. Sun, X. Wang, et al., “Foundation model enhanced mul- tiple access in heterogeneous networks,”IEEE Trans. Mobile Comput., vol. 24, no. 9, pp. 8974–8987, Sep. 2025

    2025

  3. [3]

    Toward joint learning of optimal MAC signaling and wireless channel access,

    A. Valcarce and J. Hoydis, “Toward joint learning of optimal MAC signaling and wireless channel access,”IEEE Trans. Cognit. Commun. Netw., vol. 7, no. 4, pp. 1233–1243, Oct. 2021

    2021

  4. [4]

    The emergence of wireless MAC protocols with multi-agent reinforcement learning,

    M. P. Mota, A. Valcarce, J.-M. Gorce, et al., “The emergence of wireless MAC protocols with multi-agent reinforcement learning,” inProc. IEEE Global Commun. Conf., Madrid, Spain, Dec. 2021

    2021

  5. [5]

    Scalable joint learning of wireless multiple-access policies and their signaling,

    M. P. Mota, A. Valcarce, and J.-M. Gorce, “Scalable joint learning of wireless multiple-access policies and their signaling,” inProc. IEEE Veh. Technol. Conf., Helsinki, Finland, Jun. 2022

    2022

  6. [6]

    Intent-aware DRL-based NOMA uplink dynamic scheduler for IIoT,

    S. Mostafa, M. P. Mota, A. Valcarce, et al., “Intent-aware DRL-based NOMA uplink dynamic scheduler for IIoT,” 2025. [Online]. Available: https://arxiv.org/abs/2403.18364

    arXiv 2025

  7. [7]

    Optimizing wireless discontinuous reception via MAC signaling learning,

    A. Pastore, A. A. de Dios, and ´Alvaro Valcarce, “Optimizing wireless discontinuous reception via MAC signaling learning,” 2024. [Online]. Available: https://arxiv.org/abs/2406.13834

    arXiv 2024

  8. [8]

    On learning generalized wireless MAC communication protocols via a feasible multi-agent reinforcement learning framework,

    L. Miuccio, S. Riolo, S. Samarakoon, et al., “On learning generalized wireless MAC communication protocols via a feasible multi-agent reinforcement learning framework,”IEEE Trans. Mach. Learn. Commun. Networking, vol. 2, pp. 298–317, Feb. 2024

    2024

Show all 58 references
  1. [9]

    Learning random access schemes for massive machine-type communication with marl,

    M. A. Jadoon, A. Pastore, M. Navarro, et al., “Learning random access schemes for massive machine-type communication with marl,”IEEE Trans. Mach. Learn. Commun. Netw., vol. 2, pp. 95–109, Oct. 2024

    2024

  2. [10]

    LLM4MAC: An LLM-driven reinforcement learning framework for MAC protocol emergence,

    R. Tan, R. Li, and Z. Zhao, “LLM4MAC: An LLM-driven reinforcement learning framework for MAC protocol emergence,” inProc. IEEE SPAWC, Surrey, UK, Jul. 2025

    2025

  3. [11]

    A dynamic hierarchical game approach for user association and resource allocation in hetnets with wireless backhaul,

    B. Huang and A. Guo, “A dynamic hierarchical game approach for user association and resource allocation in hetnets with wireless backhaul,” IEEE Wirel. Commun. Lett., vol. 13, no. 1, pp. 59–63, Jan. 2024

    2024

  4. [12]

    Make smart decisions faster: Deciding D2D resource allocation via Stackelberg game guided multi-agent deep reinforcement learning,

    D. Shi, L. Li, T. Ohtsuki, et al., “Make smart decisions faster: Deciding D2D resource allocation via Stackelberg game guided multi-agent deep reinforcement learning,”IEEE Trans. Mobile Comput., vol. 21, no. 12, pp. 4426–4438, Dec. 2022

    2022

  5. [13]

    Towards semantic communication protocols for 6G: From protocol learning to language-oriented approaches,

    J. Park, S.-W. Ko, J. Choi, et al., “Towards semantic communication protocols for 6G: From protocol learning to language-oriented approaches,” 2023. [Online]. Available: https://arxiv.org/abs/2310.09506

    2023 arXiv

  6. [14]

    WiLLM: an open framework for LLM services over wireless systems,

    B. Liu, Y . Lu, J. Zhao, et al., “WiLLM: an open framework for LLM services over wireless systems,” 2025. [Online]. Available: https://arxiv.org/abs/2506.19030

    2025 arXiv

  7. [15]

    Large language models for wireless communications: From adaptation to autonomy,

    L. Liang, H. Ye, Y . Sheng, et al., “Large language models for wireless communications: From adaptation to autonomy,” 2025. [Online]. Available: https://arxiv.org/abs/2507.21524

    2025

  8. [16]

    Grounding large language models in interactive environments with online reinforcement learning,

    T. Carta, C. Romac, T. Wolf, et al., “Grounding large language models in interactive environments with online reinforcement learning,” in Proc. Int. Conf. Mach. Learn., Honolulu, Hawaii, USA, Jul. 2023, p. 3676–3713

    2023

  9. [17]

    On scaling latency-aware MAC communication protocols with a hierarchical network topology,

    S. Sarbu, M. P. Mota, and M. Bennis, “On scaling latency-aware MAC communication protocols with a hierarchical network topology,” inProc. IEEE Int. Conf. Commun., Denver, CO, USA, Jun. 2024, pp. 2555–2560

    2024

  10. [18]

    ProtocolLLM: RTL benchmark for systemverilog generation of communication protocols,

    A. Sheth, I. Sheth, and M. Fritz, “ProtocolLLM: RTL benchmark for systemverilog generation of communication protocols,” 2025. [Online]. Available: https://arxiv.org/abs/2506.07945

    2025 arXiv

  11. [19]

    LLM-based emulation of the radio resource control layer: Towards AI-native RAN protocols,

    Z. Liu, B. Liu, A. Valcarce, et al., “LLM-based emulation of the radio resource control layer: Towards AI-native RAN protocols,” 2025. [Online]. Available: https://arxiv.org/abs/2505.16821

    2025

  12. [20]

    Multi-task reinforcement learning-based multiple access for dynamic wireless networks,

    Z. Chen, X. Sun, Y . Jin, et al., “Multi-task reinforcement learning-based multiple access for dynamic wireless networks,”IEEE Trans. Mobile Comput., vol. 24, no. 9, pp. 9153–9167, Sep. 2025

    2025

  13. [21]

    Knowledge distillation from language- oriented to emergent communication for multi-agent remote control,

    Y . Kim, S. Seo, J. Park, et al., “Knowledge distillation from language- oriented to emergent communication for multi-agent remote control,” in Proc. IEEE Int. Conf. Commun., vol. 23, no. 7, Denver, CO, USA, Jun. 2024, pp. 7000–7012

    2024

  14. [22]

    Toward semantic communication pro- tocols: A probabilistic logic perspective,

    S. Seo, J. Park, S.-W. Ko, et al., “Toward semantic communication pro- tocols: A probabilistic logic perspective,”IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2670–2686, Aug. 2023

    2023

  15. [23]

    CP-AgentNet: Autonomous and explainable communication protocol design using generative agents,

    D. C. Kwon and X. Zhang, “CP-AgentNet: Autonomous and explainable communication protocol design using generative agents,”

  16. [24]

    Resilient LLM-empowered semantic MAC protocols via zero-shot adaptation and knowledge distillation,

    Y . Kim, J. Park, M. Bennis, et al., “Resilient LLM-empowered semantic MAC protocols via zero-shot adaptation and knowledge distillation,”

  17. [25]

    ML framework for wireless MAC protocol design,

    N. Keshtiarast and M. Petrova, “ML framework for wireless MAC protocol design,” inProc. IEEE Int. Conf. Mach. Learn. Commun. Netw., Stockholm, Sweden, May 2024, pp. 560–565

    2024

  18. [26]

    Available: https://arxiv.org/abs/2505.21518

    [Online]. Available: https://arxiv.org/abs/2505.21518

    arXiv

  19. [27]

    Toward specialized wireless networks using an ML-driven radio interface,

    K. Szczech, M. Wojnar, K. Kosek-Szott, et al., “Toward specialized wireless networks using an ML-driven radio interface,”IEEE Access, vol. 13, pp. 141 814–141 831, Aug. 2025

    2025

  20. [28]

    Wireless MAC protocol synthesis and optimization with multi-agent distributed reinforcement learning,

    N. Keshtiarast, O. Renaldi, and M. Petrova, “Wireless MAC protocol synthesis and optimization with multi-agent distributed reinforcement learning,”IEEE Netw. Lett., vol. 6, no. 4, pp. 242–246, Dec. 2024

    2024

  21. [29]

    NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,

    Y . Chen, R. Li, Z. Zhao, et al., “NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,” IEEE Network, vol. 38, no. 6, pp. 404–413, Nov. 2024

    2024

  22. [30]

    Multi-agent reinforcement learning- based distributed channel access for next generation wireless networks,

    Z. Guo, Z. Chen, P. Liu, et al., “Multi-agent reinforcement learning- based distributed channel access for next generation wireless networks,” IEEE J. Sel. Areas Commun., vol. 40, no. 5, pp. 1587–1599, May 2022

    2022

  23. [31]

    WirelessLLM: Empowering large language models towards wireless intelligence,

    J. Shao, J. Tong, Q. Wu, et al., “WirelessLLM: Empowering large language models towards wireless intelligence,”IEEE J. Commun. Info. Netw., vol. 9, no. 2, pp. 99–112, Jun. 2024

    2024

  24. [32]

    LLM-empowered resource allocation in wireless communications systems,

    W. Lee and J. Park, “LLM-empowered resource allocation in wireless communications systems,” 2024. [Online]. Available: https://arxiv.org/abs/2408.02944

    2024 arXiv

  25. [33]

    A Stackelberg game approach for IRS- aided WPCN multicast systems,

    L. Zhai, Y . Zou, J. Zhu, et al., “A Stackelberg game approach for IRS- aided WPCN multicast systems,”IEEE Trans. Wireles Commun., vol. 21, no. 5, pp. 3249–3262, May 2022

    2022

  26. [34]

    Resource allocation for multi-tenant network slicing: A multi-leader multi-follower Stackelberg game approach,

    T. D. Tran and L. B. Le, “Resource allocation for multi-tenant network slicing: A multi-leader multi-follower Stackelberg game approach,” IEEE Trans. Veh. Technol., vol. 69, no. 8, pp. 8886–8899, Aug. 2020

    2020

  27. [35]

    Balancing performance and cost for two-hop cooperative communications: Stackelberg game and distributed multi-agent reinforcement learning,

    Y . Geng, E. Liu, W. Ni, et al., “Balancing performance and cost for two-hop cooperative communications: Stackelberg game and distributed multi-agent reinforcement learning,”IEEE Trans. Cognit. Commun. Netw., vol. 10, no. 6, pp. 2193–2208, Dec. 2024

    2024

  28. [36]

    Two layer Stackelberg game-based resource allocation in cloud-network convergence service computing,

    T. Lyu, H. Xu, F. Liu, et al., “Two layer Stackelberg game-based resource allocation in cloud-network convergence service computing,” IEEE Trans. Cognit. Commun. Netw., vol. 10, no. 6, pp. 2412–2426, Dec. 2024

    2024

  29. [37]

    3rd generation partnership project; technical specification group radio access network; NR; physical layer procedures for data (release 18),

    3GPP, “3rd generation partnership project; technical specification group radio access network; NR; physical layer procedures for data (release 18),” 3GPP, Tech. Rep. TS 38.214 V18.7.0, 2025

    2025

  30. [38]

    3rd generation partnership project; technical specification group radio access network; NR; medium access control (MAC) protocol specification (release 18),

    3GPP, “3rd generation partnership project; technical specification group radio access network; NR; medium access control (MAC) protocol specification (release 18),” 3GPP, Tech. Rep. TS 38.321 V18.6.0, 2025

    2025

  31. [39]

    A quantitative measure of fairness and discrimination for resource allocation in shared computer systems,

    R. Jain, D. Chiu, and W. Hawe, “A quantitative measure of fairness and discrimination for resource allocation in shared computer systems,”

  32. [40]

    Hashing for similarity search: A survey,

    J. Wang, H. T. Shen, J. Song, et al., “Hashing for similarity search: A survey,” 2014

    2014

  33. [41]

    Stackelberg actor-critic: Game- theoretic reinforcement learning algorithms,

    L. Zheng, T. Fiez, Z. Alumbaugh, et al., “Stackelberg actor-critic: Game- theoretic reinforcement learning algorithms,”AAAI Conf. Artif. Intell., vol. 36, no. 8, pp. 9217–9224, Jun. 2022

    2022

  34. [42]

    Implicit learning dynamics in Stackelberg games: equilibria characterization, convergence analysis, and empirical study,

    T. Fiez, B. Chasnov, and L. Ratliff, “Implicit learning dynamics in Stackelberg games: equilibria characterization, convergence analysis, and empirical study,” inProc. Int. Conf. Mach. Learn., Virtual Edition, Jul. 2020, p. 3133–3144

    2020

  35. [43]

    Decentralized consensus inference-based hierarchical reinforcement learning for multiconstrained UA V pursuit- evasion game,

    Y . Xiang, S. Li, R. Li, et al., “Decentralized consensus inference-based hierarchical reinforcement learning for multiconstrained UA V pursuit- evasion game,”IEEE Trans. Neural Netw. Learn. Syst., 2025, early access

    2025

  36. [44]

    Learning in conjectural Stackelberg games,

    F. Morri, H. L. Cadre, and L. Brotcorne, “Learning in conjectural Stackelberg games,” 2025. [Online]. Available: https://arxiv.org/abs/2501.13686

    2025 arXiv

  37. [45]

    Actor-critic algorithms,

    V . Konda and J. Tsitsiklis, “Actor-critic algorithms,”Adv. Neural Inf. Process. Syst., p. 1008–1014, 1999

    1999

  38. [46]

    On the characterization of local Nash equilibria in continuous games,

    L. J. Ratliff, S. A. Burden, and S. S. Sastry, “On the characterization of local Nash equilibria in continuous games,”IEEE Trans. Autom. Contr., vol. 61, no. 8, pp. 2301–2307, jun. 2016

    2016

  39. [47]

    The Stone-Weierstrass theorem and its application to neural networks,

    N. Cotter, “The Stone-Weierstrass theorem and its application to neural networks,”IEEE Trans. Neural Networks, vol. 1, no. 4, pp. 290–295, dec. 1990

    1990

  40. [48]

    The Llama 3 herd of models,

    A. Grattafiori, A. Dubey, A. Jauhri, et al., “The Llama 3 herd of models,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21783

    2024 arXiv

  41. [49]

    Joint task assignment and spectrum allocation in heterogeneous UA V communication networks: A coalition formation game-theoretic approach,

    J. Chen, Q. Wu, Y . Xu, et al., “Joint task assignment and spectrum allocation in heterogeneous UA V communication networks: A coalition formation game-theoretic approach,”IEEE Trans. Wirel. Commun., vol. 20, no. 1, pp. 440–452, jan. 2021

    2021

  42. [50]

    Markov decision processes,

    M. L. Puterman, “Markov decision processes,” inStochastic Models, ser. Handbooks in Operations Research and Management Science. Elsevier, 1990, vol. 2, pp. 331–434. 13

    1990

  43. [51]

    Granas, J

    A. Granas, J. Dugundji, et al.,Fixed point theory. Springer, 2003, vol. 14

    2003

  44. [52]

    Multi-agent feedback enabled neural networks for intelligent communications,

    F. Sun, Y . Li, Y . Wen, et al., “Multi-agent feedback enabled neural networks for intelligent communications,”IEEE Trans. Wirel. Commun., vol. 21, no. 8, pp. 6167–6179, feb. 2022

    2022

  45. [53]

    Basic properties of the Schur complement,

    R. A. Horn and F. Zhang, “Basic properties of the Schur complement,” inThe Schur complement and its applications. Springer, 2005, pp. 17–46

    2005

  46. [54]

    R. A. Horn and C. R. Johnson,Topics in matrix analysis. Cambridge University Press, 1994. 14 APPENDIXA PROOF OFTHEOREM1 For ease of reference, we first present a table summarizing the notations used throughout the proofs as Table V. To establish the existence of an ESE, we fir...

    1994

  47. [55]

    J. M. Ortega and W. C. Rheinboldt,Iterative solution of nonlinear equations in several variables. SIAM, 2000

    2000

  48. [58]

    Similarly, he bottom-right blockD ι =ι uD= 1 2 (J22 +J ⊤ 22), and its(j, k)-th sub- block is denoted by ιu 2 ∇2 θj,u,θk,u J (e) j,u + (∇2 θk,u,θj,uJ (e) k,u)⊤

    =∇ 2 θb J (e) b (Θ∗), is the Hessian of the BS with respect to its strategy parameters. Similarly, he bottom-right blockD ι =ι uD= 1 2 (J22 +J ⊤ 22), and its(j, k)-th sub- block is denoted by ιu 2 ∇2 θj,u,θk,u J (e) j,u + (∇2 θk,u,θj,uJ (e) k,u)⊤ . From the SOCs of the underly...

  49. [1998]

    Available: https://arxiv.org/abs/cs/9809099

    [Online]. Available: https://arxiv.org/abs/cs/9809099

    arXiv

  50. [2025]

    Available: https://arxiv.org/abs/2503.17850

    [Online]. Available: https://arxiv.org/abs/2503.17850

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.