Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Resilient LLM-Empowered Semantic MAC Protocols via Zero-Shot Adaptation and Knowledge Distillation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an off-the-shelf LLM can act as both a first-responder medium-access controller and a teacher for a distilled neural protocol, and that switching from the LLM to the distilled model once it statistically overtakes…

desk verdict Worth engaging: the T3NPM hybrid and meta-resilience metric are real contributions, but the headline gains depend on prompt selection on the target shift and single-seed runs. read the letter →

arxiv 2505.21518 v1 pith:HTRSEKGX submitted 2025-05-22 cs.NI

Yongjun Kim , Jihong Park , Mehdi Bennis , Junil Choi This is my paper
classification cs.NI
keywords semanticMACprotocoltoken-basedcommunicationlearninglargelanguagemodelknowledgedistillationmeta-resiliencepromptoptimization
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 addresses a known brittleness of neural medium access control (MAC) models: their learned control messages collapse when the network changes, for instance when the number of user devices grows, because the neural architecture cannot accommodate the new dimension without retraining. The authors propose to use a large language model (LLM) as a MAC controller that reads buffer states in natural language and outputs actions, which works immediately after a shift with no retraining. They then distill the LLM's knowledge into a small neural protocol model, and introduce a hybrid scheme that runs the LLM first and switches to the distilled model once a statistical test shows the distilled model has become better. To compare such schemes fairly when the post-shift performance target is unknown, they define a new metric, meta-resilience, which averages resilience across all target goodput levels. Simulations show the hybrid scheme achieves the highest meta-resilience, beating a standard neural protocol by 20.56% and slotted ALOHA by 23.53%, while cutting compute by 19.8x compared to using the LLM throughout.

What carries the argument

The paper's argument is carried by three protocol models and one metric. TPM (token-based protocol model) runs an LLM as the base station: UEs report buffer occupancy in a fixed textual query, and the LLM, guided by a natural-language instruction prompt, answers with one of three actions per UE; the prompt itself is optimized by TextGrad, an LLM-based automatic prompt refiner. T2NPM (TPM-to-NPM) distills TPM into a compact neural protocol: the LLM's token probabilities for the three actions, after a masked softmax over the action vocabulary, are used as teacher logits in a KL-divergence knowledge-distillation loss combined with the usual temporal-difference loss, so the small network inherits the LLM's policy while still learning from environment interaction. T3NPM (TPM-after-T2NPM) runs TPM in the early phase after a shift and switches to T2NPM once a one-sided Mann-Whitney U test on goodput measurements rejects the null that TPM is at least as good; a mix-scheduling rule reuses past measurements so that a small measurement window per episode still yields enough samples. Finally, meta-resilience is a new evaluation metric: resilience at a fixed target goodput is the area-under-curve ratio $\min(G_n/\hat{G},1)$, and meta-resilience integrates that over all feasible target levels, which is what lets the authors compare protocols without knowing the post-shift oracle performance.

What would settle it

Measure the LLM's action-accuracy on held-out buffer states that were not used in prompt selection and compare against an oracle policy; if accuracy is near chance or the fraction of responses violating the required output format is high, then TPM's early-phase goodput advantage, and with it T3NPM's margin, collapses.

Watch

Extended reading notes

Core claim

The central discovery is that an off-the-shelf LLM can be repurposed, with only a natural-language instruction optimized by a text-based automatic prompt optimizer (TextGrad), to make acceptable MAC scheduling decisions under an environmental shift, and that this coarse policy can serve as a teacher. By extracting the LLM's token probabilities for the three actions and using them as soft targets in a knowledge-distillation loss, the small neural protocol (T2NPM) is shown to converge faster and to a higher goodput than the same architecture trained with reinforcement learning alone (NPM). Combining the two roles in T3NPM, with a Mann-Whitney U test deciding when T2NPM statistically outperforms TPM and a mix-switching rule reusing past measurements to save measurement time, yields the highest meta-resilience across all compared schemes. The paper's own claim, stated in Sec. VI, is the numerical margin: T3NPM exceeds T2NPM, TPM, NPM, and S-ALOHA by 4.37%, 10.05%, 20.56%, and 23.53% in meta-resilience, with a 19.8x reduction in FLOPS relative to TPM.

Load-bearing premise

The load-bearing premise is that an off-the-shelf large language model, given only a written set of rules and the current buffer sizes of the user devices, can immediately choose transmit/delete/wait actions well enough to outperform the collapsed neural protocol, even though the paper never measures how accurately the model follows its instructions.

Editorial extensions

If this is right

  • If T3NPM is correct, a network operator facing a sudden increase in devices can keep MAC goodput stable by switching to an LLM-driven controller immediately, without waiting for retraining.
  • The same LLM can double as a teacher, so the retraining phase recovers faster and converges higher than retraining from scratch, and the finally deployed model is a small network rather than a multi-billion-parameter LLM.
  • Computational cost during the recovery period can be cut by a factor of 19.8 while still improving resilience, making LLM-assisted recovery feasible even where full LLM inference per time slot is too expensive.
  • Because meta-resilience integrates over all target goodput levels, it provides a single number for comparing protocols under unknown post-shift performance and could be adopted as a standard evaluation metric for shift-resilient wireless designs.

Reading between the lines

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

  • An extension left implicit in the paper: the same teacher-first, student-after switch could apply to other resource-allocation tasks where a general model gives an immediate but coarse policy and a small student can be distilled from it, though the reported margins are likely dependent on the specific task and LLM.
  • A testable extension would replace the paper's 10.7-billion-parameter LLM with a smaller or larger model to see whether the early-phase advantage scales with model capability and whether the measurement window needs to grow when the LLM becomes less reliable.
  • The paper's meta-resilience metric depends on the chosen range of target goodputs, so an implicit consequence is that comparing two protocols on a narrow range near the operating point could order them differently than the integrated metric does.
  • The authors do not measure how accurately the LLM follows its instructions; a natural next step is to test whether format violations are the main cause of TPM's goodput ceiling, which would suggest constrained decoding or stricter prompt formatting as a follow-up.
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 three LLM-empowered semantic MAC protocol frameworks for resilience under environmental shifts (focusing on an increase in the number of UEs). TPM uses a large language model as a token-based MAC controller whose instruction is refined via TextGrad; T2NPM distills TPM knowledge into a small neural protocol via a KD loss; T3NPM starts with TPM and switches to T2NPM once a Mann-Whitney U test indicates that T2NPM has become statistically better. A new meta-resilience metric is introduced to compare protocols when the target performance level after a shift is unknown. Simulations report that T3NPM achieves the highest meta-resilience, outperforming T2NPM, TPM, NPM, and S-ALOHA by 4.37%, 10.05%, 20.56%, and 23.53%, respectively, with 19.8× lower FLOPS than TPM.

Significance. If the results hold, the paper provides a useful blueprint for using an off-the-shelf LLM as both a first-responder MAC controller and a teacher for distilling a lightweight neural protocol after an environmental shift. The proposed meta-resilience metric addresses a real need for comparing protocols when the post-shift target goodput is unknown, and the compute-reduction comparison in Table II is a concrete practical contribution. The paper is clearly written and the protocol math, including the KD formulation and the switching algorithm, is mostly coherent. The central weakness is that the final prompt is selected on the same shift used for evaluation, so the claimed "zero-shot" adaptation and the headline 20.56% meta-resilience margin are not yet demonstrated in an externally valid way.

major comments (4)
  1. [III-B, VI-C, VI-D] The claim that TPM provides zero-shot adaptation is not supported by the evaluation protocol. Section III-B selects the final instruction as φ = arg max_{φ_m} G_{φ_m}, and Fig. 10a measures G_{φ_m} under the same L:2→3 shift that is later used for the T3NPM comparison in Fig. 9. Thus TPM's instruction is effectively optimized on the target evaluation shift, so the stable early goodput of TPM—and the 20.56% meta-resilience margin over NPM—may be an artifact of test-set prompt selection rather than evidence of zero-shot generalization. The paper should evaluate φ7 on a held-out shift (e.g., L:2→4 or a different pa/bmax change) without further prompt optimization, or should rephrase the claim as one-shot prompt selection on the target shift. Additionally, no instruction-following accuracy metric is reported, so the load-bearing premise that SOLAR-10.7B can reliably parse buffer counts and output correct actions is unverified.
  2. [V-A, Eq. (15)] The definition of meta-resilience in Eq. (15) is an integral over an unspecified range of target goodput values. As written, the integral over (0,∞) diverges for any positive G_n, because ∫ min(G_n/g,1) dg = G_n + ∫_{G_n}^∞ G_n/g dg is unbounded. The text says that integration is over "a predetermined feasible range" and the numerics in Fig. 9b use [0.01,1], but the equation itself does not state these limits. The authors should either specify finite integration limits in Eq. (15) or justify the chosen range; otherwise the reported meta-resilience values are not well defined by the stated metric.
  3. [IV-A, Eq. (10)] Eq. (10) constructs the teacher's soft targets by applying a temperature-scaled softmax to token generation probabilities, not to logits. In standard knowledge distillation, the softmax is applied to logits before normalization; re-normalizing already-normalized probabilities with a softmax changes the distribution in a way that depends on the probabilities' absolute values and is not the intended use of the temperature parameter. This issue affects the teacher knowledge used in LKD (Eq. (11)) and therefore the operation of T2NPM. Please either extract the teacher's action distribution from the LLM's logits directly, or justify mathematically why applying softmax to probabilities is appropriate.
  4. [VI] All reported results appear to come from a single simulation run, with no seeds, error bars, or confidence intervals. Given that TextGrad is stochastic (Fig. 10a shows non-monotonic behavior across epochs) and the switching decision relies on small samples in the Mann-Whitney U test, the claimed performance differences of 4.37%–23.53% between protocols may plausibly lie within run-to-run variance. The paper should report multiple seeds and statistical significance of the meta-resilience comparisons before the headline claims can be accepted.
minor comments (5)
  1. [V-B, Eqs. (16)-(17)] The Mann-Whitney U test is described as testing the medians of V_TPM and V_T2NPM, but the test actually assesses stochastic dominance under the usual assumption of identically shaped distributions; please correct the hypothesis wording or use a test that directly targets medians.
  2. [Table I] The circled symbols 1© and 2© in Table I are not defined in the table or caption; please clarify what these markers denote.
  3. [Algorithm 1, line 7] The line "Generate V~T2NPM_n through Θ_n within T_M TTIs" should specify how many goodput samples are collected per episode and how the requirement of |V_T2NPM_n| = 12 is satisfied for the test.
  4. [VI-D, Fig. 10a] The "random" baseline in Fig. 10a is not fully described; please clarify how many random instructions were sampled and whether the best random instruction or the average over random instructions is reported.
  5. [V-B] There is a typo in the sentence "following the procedure in [36], , such that p≥ α"; it should read something like "following the procedure in [36], if the p-value satisfies p≥α, the null hypothesis is accepted."

Circularity Check

2 steps flagged · score 6.0 of 10

TPM's zero-shot goodput is a test-set prompt maximum, and T3NPM's claimed margins over T2NPM and TPM are by construction of the TM optimization; the NPM/S-ALOHA comparison retains independent empirical content.

  1. fitted input called prediction [Sec. III-B (TextGrad-aided TPM); Sec. VI-D, Fig. 10a; Sec. VI-C, Fig. 9]
    "The final instruction φ is selected to be the instruction φm that achieves the highest goodput across TPM operations as φ = arg max φm Gφm. ... At each epoch m, Gφm is evaluated using TPM with instruction φ = φm. ... the highest TPM goodput is achieved among all evaluated instructions using φ = φ7. ... instruction φ = φ7 is adopted for all simulations, including those in Fig. 9."

    The prompt φ7 was selected by maximizing TPM goodput on the same L↑ shift that is later used for the headline meta-resilience comparison in Fig. 9. The paper's claim of zero-shot/immediate TPM adaptation is therefore not independently predicted: the reported early-phase TPM goodput is the argmax over ten candidate prompts on the target environment. The assertion that the optimized instruction is environment-independent is not tested by any held-out shift, so the TPM and T3NPM advantages that depend on TPM's stable goodput are partly fitted values rather than out-of-sample predictions.

  2. self definitional [Sec. V-B, Eq. (18); Sec. VI-F, Fig. 11b; Sec. VI-C, Fig. 9c]
    "When TM = 0 (i.e., no measurement) or TM = T (i.e., no re-training), T3NPM becomes equivalent to T2NPM or TPM, respectively. ... According to Fig. 11b, the configuration TM = 24 achieves the highest meta-resilience ... Consequently, TM = 24 is selected as the optimal solution to (18), and this configuration is adopted for all simulations, including Fig. 9. ... T3NPM's ar R({G_n}|24,5) achieves the highest meta-resilience, outperforming T2NPM, TPM, NPM, and S-ALOHA by 4.37%, 10.05%, 20.56%, and 23.53%, respectively."

    T2NPM and TPM are exactly the boundary cases TM=0 and TM=T of the T3NPM family, and TM=24 is chosen as the argmax of the same meta-resilience metric over a grid that includes those endpoints. Consequently, the conclusion that T3NPM outperforms T2NPM by 4.37% and TPM by 10.05% is true by construction of the optimization: these are the differences between the selected maximum and the boundary values of the very curve being maximized. This part of the headline is not an independent empirical discovery, although the 20.56% margin over NPM and 23.53% margin over S-ALOHA still retain genuine simulation content.

full rationale

The paper contains no load-bearing self-citation chain, no imported uniqueness theorem, and no renaming of a known result as new theory; the NPM baseline, TextGrad, SOLAR-10.7B, and the Mann-Whitney U test are all externally grounded. The circularity score comes instead from two places where a reported result is forced by how the experiment was set up. First, the TPM instruction used in the main evaluation is selected by goodput on the same environmental shift used for that evaluation, so TPM's 'zero-shot' early goodput is a selected maximum rather than an independent prediction; no held-out shift or instruction-following metric is provided to establish that φ7 encodes general MAC rules. Second, the T3NPM switching parameter TM is optimized against the same meta-resilience metric later used for comparison, and T2NPM and TPM are the endpoints TM=0 and TM=T of that same optimization family; hence the claimed advantages over T2NPM and TPM are definitional consequences of choosing the best point on the curve. The comparison against NPM and S-ALOHA is not circular in the same way, and the goodput trajectories in Fig. 9 do provide independent support for the general protocol-learning idea, but the size of the headline margin is inflated by test-set prompt selection and by reporting the optimized configuration against its own boundary cases. Overall, the central derivation is partial: the framework itself is not derived from its own conclusion, but two load-bearing numerical claims reduce by construction.

Assumptions & free parameters 10 free parameters · 6 assumptions · 1 invented entities

The central claim relies on a set of hand-chosen RL hyperparameters, a selected prompt, and an implicitly bounded metric; the free parameters above are the main ones that are not externally fixed. The axioms capture the domain assumptions about LLM capability and error-free control channels. No physical entities are invented; the meta-resilience metric is included as a conceptual construct because its definition directly affects the reported outcome.

free parameters (10)
  • Reward shaping constants = rho1=10, rho2=8, rho3=rho4=4, rho5=1
    Hand-chosen in Sec. VI-A to shape DQN training; no sensitivity analysis.
  • DQN exploration schedule = initial 1, decay 0.9, floor 0.1
    Epsilon-greedy schedule in Sec. VI-A; standard but tuned.
  • TD discount and soft update = gamma=0.99, sigma=1e-3
    Sec. VI-A; common DQN hyperparameters.
  • KD temperature = kappa=2
    Sec. VI-A; affects softness of teacher/student distributions; no ablation.
  • KD loss weights = lambda1=0.1, lambda2=0.9
    Eq. (13); chosen values heavily favor KD over TD loss; no sensitivity study.
  • Measurement and MixSwitch window = TM=24, k=5
    Selected to maximize meta-resilience in Eq. (18); this is fitting on the evaluation metric.
  • Hypothesis test threshold = alpha=0.05
    One-sided Mann-Whitney U threshold in Sec. VI-A; conventional but arbitrary.
  • Meta-resilience integration range = [0.01,1]
    Appears only in Sec. VI-C figures; Eq. (15) is unbounded, so the metric's value depends on this implicit choice.
  • S-ALOHA transmission probability = 0.33
    Fixed classical baseline in Sec. VI-A.
  • TextGrad instruction = phi_7
    Chosen as argmax goodput G_phi_m in Sec. III-B; in-sample selection on the evaluation scenario.
assumptions (6)
  • domain assumption The LLM's pretrained knowledge suffices to act as an immediate MAC scheduler from natural-language buffer states.
    Sec. III-A and Fig. 5; if this fails, TPM and T3NPM lose their early-phase advantage.
  • domain assumption Control channels for UCM/DCM are error-free.
    Sec. II-A states 'for simplicity and without loss of generality'; this removes control-message loss from the resilience analysis.
  • domain assumption LLM token generation probabilities at the action token correlate with action correctness and can serve as teacher logits.
    Sec. IV-A, based on [26]; Eq. (10) transforms these probabilities into teacher targets for KD.
  • ad hoc to paper Meta-resilience over a finite feasible target range is the right objective for comparing protocols after unknown shifts.
    Sec. V-A Eq. (15); the range is not justified and the metric is introduced by the same paper that is then optimized on it.
  • domain assumption A single environmental shift, L going from 2 to 3, is representative enough to validate resilience claims.
    Sec. VI-B justifies focusing on L up arrow after Table I, but all headline T3NPM numbers come from this single shift.
  • standard math Standard DQN/TD learning with target network converges as described.
    Sec. II-C Eq. (4); standard RL machinery with no formal convergence proof.
invented entities (1)
  • Meta-resilience metric Rbar({G_n})
    purpose: Quantifies resilience without a fixed target goodput by averaging resilience over a range of targets (Eq. 15).
    Introduced and used by the same paper; integration range is implicit and no external benchmark validates it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Resilient LLM-Empowered Semantic MAC Protocols via Zero-Shot Adaptation and Knowledge Distillation." pith.science (2026). https://pith.science/paper/HTRSEKGX

@misc{pith2026250521518,
  author       = {Pith},
  title        = {Pith review of: Resilient LLM-Empowered Semantic MAC Protocols via Zero-Shot Adaptation and Knowledge Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTRSEKGX}},
  note         = {Machine review of arXiv:2505.21518}
}
read the original abstract

Neural network-based medium access control (MAC) protocol models (NPMs) improve goodput through site-specific operations but are vulnerable to shifts from their training network environments, such as changes in the number of user equipments (UEs) severely degrade goodput. To enhance resilience against such environmental shifts, we propose three novel semantic MAC protocol frameworks empowered by large language models (LLMs). First, we introduce a token-based protocol model (TPM), where an LLM generates MAC signaling messages. By editing LLM instruction prompts, TPM enables instant adaptation, which can be further enhanced by TextGrad, an LLM-based automated prompt optimizer. TPM inference is fast but coarse due to the lack of real interactions with the changed environment, and computationally intensive due to the large size of the LLM. To improve goodput and computation efficiency, we develop T2NPM, which transfers and augments TPM knowledge into an NPM via knowledge distillation (KD). Integrating TPM and T2NPM, we propose T3NPM, which employs TPM in the early phase and switches to T2NPM later. To optimize this phase switching, we design a novel metric of meta-resilience, which quantifies resilience to unknown target goodput after environmental shifts. Simulations corroborate that T3NPM achieves 20.56% higher meta-resilience than NPM with 19.8x lower computation cost than TPM in FLOPS.

Figures

Figures reproduced from arXiv: 2505.21518 by the authors.

Figure 1
Figure 1. LLM-empowered semantic MAC protocols. as the neural protocol model (NPM) [2]. NPM is built on multi-agent deep reinforcement learning (MADRL), wherein user equipments (UEs) are mapped into multiple agents that communicate control messages through a set of shared neu￾ral network (NN) layers associated with their base station (BS). By training the MADRL in specific environments and applications, NPM dynamically learns… view at source ↗
Figure 2
Figure 2. Network model under environmental shifts. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. A schematic illustration of TPM. can be reused across different environments, as it describes the task itself rather than the environment-specific parameters. A. TPM Construction and Operation Unlike the uninterpretable control messages in NPM, the UCMs and DCMs in TPM are expressed in natural language, based on the architecture depicted in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: TextGrad example at epoch m. instructions that remain valid even after environmental shifts. While prior work in prompt engineering has focused largely on heuristic, handcrafted instruction design [21], TextGrad [22] enables automatic prompt refinement using an LLM. Th…
Figure 8
Figure 8. Figure 8: Episode temporal structure, comprising measuremen [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: Resilience vs. meta-resilience. elaborate in the following subsections, this metric enables both fair performance comparisons across various protocol models and optimization of T3NPM’s transition timing. A. Meta-resilience Since our proposed methods focus on environmen…
Figure 9
Figure 9. Figure 9: Goodput, resilience, and meta-resilience of semant [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Impact of TextGrad on TPM and T2NPM with respect to th [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 12
Figure 12. Figure 12: Impact of SNR on resilience. G. Ablation Studies: Impact of ICL and SNR As discussed in Sec. III-B, ICL and prompt engineering via TextGrad are two promising LLM adaptation methods for TPM construction. However, ICL is effective only under mild environmental shifts, a…
Figure 11
Figure 11. Figure 11: Resilience and meta-resilience for different [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Sign in 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. LLM-Empowered Agentic MAC Protocols: A Dynamic Stackelberg Game Approach

    cs.AI 2025-10 conditional novelty 5.0 of 10

    An LLM-based multi-agent reinforcement learning framework that models uplink MAC scheduling as a Stackelberg game reports 77.6% higher throughput and 65.2% better fairness in simulation.

  2. Resilient-native and Intelligent NextG Systems

    cs.ET 2025-06 unverdicted novelty 3.0 of 10

    A position paper defining resilience for 6G networks as resisting, recovering from, and adapting to unforeseen disruptions, with a proposed multi-disciplinary mathematical toolkit.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages · cited by 2 Pith papers

  1. [1]

    T he Role of AI in 6G MAC,

    A. V alcarce, P . Kela, S. Mandelli, and H. Viswanathan, “T he Role of AI in 6G MAC,” in 2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit) . IEEE, Jul 2024, pp. 723–728

    work page 2024

  2. [2]

    The Emerge nce of Wireless MAC Protocols with Multi-Agent Reinforcement Lea rning,

    M. Mota, A. V alcarce, J. Gorce, and J. Hoydis, “The Emerge nce of Wireless MAC Protocols with Multi-Agent Reinforcement Lea rning,” IEEE Globecom W orkshops (GC Wkshps) , Jan 2021

    work page 2021

  3. [3]

    Towa rd Semantic Communication Protocols: A Probabilistic Logic P erspective,

    S. Seo, J. Park, S. Ko, J. Choi, M. Bennis, and S. Kim, “Towa rd Semantic Communication Protocols: A Probabilistic Logic P erspective,” IEEE Journal on Selected Areas in Communications , vol. 41, pp. 2670 – 2686, Jun 2023

    work page 2023

  4. [4]

    Evolution of MAC Protocols in the Machine Learning Decade: A Comprehensive Survey

    M. Hussien et al. , “Evolution of MAC Protocols in the Ma- chine Learning Decade: A Comprehensive Survey,” arXiv preprint arXiv:2302.13876, Jan 2023

    work page Pith review arXiv 2023

  5. [5]

    A Deep-Learni ng-based MAC for Integrating Channel Access, Rate Adaptation, and Ch annel Switch,

    J. Xin, W. Xu, B. Cao, T. Wang, and S. Zhang, “A Deep-Learni ng-based MAC for Integrating Channel Access, Rate Adaptation, and Ch annel Switch,” Digital Communications and Networks , Oct 2024

    work page 2024

  6. [6]

    Wireless MA C Protocol Synthesis and Optimization with Multi-Agent Distributed R einforcement Learning,

    N. Keshtiarast, O. Renaldi, and M. Petrova, “Wireless MA C Protocol Synthesis and Optimization with Multi-Agent Distributed R einforcement Learning,” IEEE Networking Letters , Nov 2024

    work page 2024

  7. [7]

    Learning MAC Protocols in HetNets: A Cooperative Multi-Agent Deep Reinf orcement Learning Approach,

    F. Naeem, N. Adam, G. Kaddoum, and O. Waqar, “Learning MAC Protocols in HetNets: A Cooperative Multi-Agent Deep Reinf orcement Learning Approach,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC) . IEEE, Jul 2024, pp. 1–6

    work page 2024

  8. [8]

    Large AI Model Empowered Multimodal Semantic Communications,

    F. Jiang et al. , “Large AI Model Empowered Multimodal Semantic Communications,” IEEE Communications Magazine , Sep 2024

    work page 2024

Show all 38 references
  1. [9]

    Large Language Models for Telecom: The Next Big Thing?

    L. Bariah, Q. Zhao, H. Zou, Y . Tian, F. Bader, and M. Debbah , “Large Language Models for Telecom: The Next Big Thing?” arXiv preprint arXiv:2306.10249, 2023

    2023 arXiv

  2. [10]

    Beam Prediction based on Large Language Models,

    Y . Sheng et al. , “Beam Prediction based on Large Language Models,” IEEE Wireless Communications Letters , Feb 2025

    2025

  3. [11]

    Port -LLM: A Port Prediction Method for Fluid Antenna based on Large Lang uage Models,

    Y . Zhang, H. Yin, W. Li, E. Bjornson, and M. Debbah, “Port -LLM: A Port Prediction Method for Fluid Antenna based on Large Lang uage Models,” arXiv preprint arXiv:2502.09857 , Feb 2025

    2025 arXiv

  4. [12]

    LLM-Augmented Deep R einforce- ment Learning for Dynamic O-RAN Network Slicing,

    F. Lotfi, H. Rajoli, and F. Afghah, “LLM-Augmented Deep R einforce- ment Learning for Dynamic O-RAN Network Slicing,” environments, vol. 2, p. 4

  5. [13]

    LLM-xApp: A Large Language Model Empowered Radio Resource Management xApp fo r 5G O-RAN,

    X. Wu, J. Farooq, Y . Wang, and J. Chen, “LLM-xApp: A Large Language Model Empowered Radio Resource Management xApp fo r 5G O-RAN,” in Symposium on Networks and Distributed Systems Security (NDSS), W orkshop on Security and Privacy of Next-G eneration Networks (FutureG 2025), ...

    2025

  6. [14]

    Towards Secure Intelligent O-RAN Architec- ture: Vulnerabilities, Threats and Promising Technical So lutions using LLMs,

    M. K. Motalleb et al. , “Towards Secure Intelligent O-RAN Architec- ture: Vulnerabilities, Threats and Promising Technical So lutions using LLMs,” arXiv preprint arXiv:2411.08640 , Nov 2024

    2024 arXiv

  7. [15]

    LLM4MAC: An LLM-Driven Rein- forcement Learning Framework for MAC Protocol Emergence,

    R. Tan, R. Li, and Z. Zhao, “LLM4MAC: An LLM-Driven Rein- forcement Learning Framework for MAC Protocol Emergence,” arXiv preprint arXiv:2503.08123, Mar 2025

    2025 arXiv

  8. [16]

    Multiagent Cooperation and Competition with Deep Reinforcement Learning,

    A. Tampuu et al. , “Multiagent Cooperation and Competition with Deep Reinforcement Learning,” PLOS ONE 12 (4):e0172395. , Apr 2017

    2017

  9. [17]

    R. S. Sutton, A. G. Barto et al. , Reinforcement Learning: An Introduc- tion. MIT press Cambridge, Mar 1999, vol. 1, no. 1

    1999

  10. [18]

    Playing Atari with Deep Reinforcement Learni ng,

    V . Mnih, “Playing Atari with Deep Reinforcement Learni ng,” arXiv preprint arXiv:1312.5602, Dec 2013

    2013 arXiv

  11. [19]

    Learning to Predict by the Methods of Temp oral Differ- ences,

    R. S. Sutton, “Learning to Predict by the Methods of Temp oral Differ- ences,” Machine learning , vol. 3, pp. 9–44, Aug 1988

    1988

  12. [20]

    Continuous Control with Deep Reinforce ment Learning,

    T. Lillicrap, “Continuous Control with Deep Reinforce ment Learning,” arXiv preprint arXiv:1509.02971 , Jul 2019

    2019 arXiv

  13. [21]

    Toward s Hierarchical Multi-Agent Workflows for Zero-Shot Prompt Optimization,

    Y . Liu, J. Singh, G. Liu, A. Payani, and L. Zheng, “Toward s Hierarchical Multi-Agent Workflows for Zero-Shot Prompt Optimization,” arXiv preprint arXiv:2405.20252, Apr 2025

    2025 arXiv

  14. [22]

    TextGrad: Automatic “Differentiation

    M. Y uksekgonul et al. , “TextGrad: Automatic “Differentiation” via Text,” arXiv preprint arXiv:2406.07496 , Jun 2024

    2024 arXiv

  15. [23]

    Understanding Telecom Language Through Large Language Mo dels,

    L. Bariah, H. Zou, Q. Zhao, B. Mouhouche, F. Bader, and M. Debbah, “Understanding Telecom Language Through Large Language Mo dels,” in 2023 IEEE Global Communications Conference . IEEE, Feb 2023, pp. 6542–6547

    2023

  16. [24]

    Prompt Engineering or Fine Tuning: An Empirical Assessment of Large Language Models in Automated Software E ngi- neering Tasks,

    J. Shin et al. , “Prompt Engineering or Fine Tuning: An Empirical Assessment of Large Language Models in Automated Software E ngi- neering Tasks,” arXiv preprint arXiv:2310.10508 , Feb 2025

    2025 arXiv

  17. [25]

    Towards Understanding Knowl edge Distil- lation,

    M. Phuong and C. Lampert, “Towards Understanding Knowl edge Distil- lation,” International conference on machine learning , pp. 5142–5151, Jun 2019

    2019

  18. [26]

    Softmax Probabiliti es (Mostly) Predict Large Language Model Correctness on Multiple-Choi ce Q&A,

    B. Plaut, K. Nguyen, and T. Trinh, “Softmax Probabiliti es (Mostly) Predict Large Language Model Correctness on Multiple-Choi ce Q&A,” arXiv:2402.13213, Feb 2024

    2024 arXiv

  19. [27]

    Born Again Neural Networks,

    T. Furlanello et al. , “Born Again Neural Networks,” in International conference on machine learning . PMLR, Jul 2018, pp. 1607–1616

    2018

  20. [28]

    Design and Assessment Method ology for System Resilience Metrics,

    M. Najarian and G. J. Lim, “Design and Assessment Method ology for System Resilience Metrics,” Risk Analysis , vol. 39, no. 9, pp. 1885– 1898, Feb 2019

    2019

  21. [29]

    Modelling Communication Network Challenges for Future Internet Resilience, Survivability, and Disrup tion Tolerance: a Simulation-based Approach,

    E. K. Cetinkaya et al., “Modelling Communication Network Challenges for Future Internet Resilience, Survivability, and Disrup tion Tolerance: a Simulation-based Approach,” Telecommunication Systems, vol. 52, pp. 751–766, Sep 2013

    2013

  22. [30]

    Methodology for Assessing the Resilience of Networked Infrastructure,

    D. A. Reed, K. C. Kapur, and R. D. Christie, “Methodology for Assessing the Resilience of Networked Infrastructure,” IEEE Systems Journal, vol. 3, pp. 174 – 180, May 2009

    2009

  23. [31]

    Come back Kid: Resilience for Mixed-Critical Wireless Network Resource M anagement,

    R.-J. Reifert, S. Roth, A. A. Ahmad, and A. Sezgin, “Come back Kid: Resilience for Mixed-Critical Wireless Network Resource M anagement,” IEEE Transactions on V ehicular Technology, Jul 2023

    2023

  24. [32]

    A Framework to Quantitatively Assess and Enhance the Seismic Resilience of Communities,

    M. Bruneau et al., “A Framework to Quantitatively Assess and Enhance the Seismic Resilience of Communities,” Earthquake spectra , vol. 19, no. 4, pp. 733–752, Nov 2003

    2003

  25. [33]

    Model-Agnostic Meta -Learning for Fast Adaptation of Deep Networks,

    C. Finn, P . Abbeel, and S. Levine, “Model-Agnostic Meta -Learning for Fast Adaptation of Deep Networks,” in International conference on machine learning. PMLR, Aug 2017, pp. 1126–1135

    2017

  26. [34]

    The Meta Distribution of the SIR in Poisson Bipolar and Cellular Networks,

    M. Haenggi, “The Meta Distribution of the SIR in Poisson Bipolar and Cellular Networks,” IEEE Transactions on Wireless Communications , vol. 15, no. 4, pp. 2577–2589, Dec 2015

    2015

  27. [35]

    Why Comparing Single Perfo rmance Scores Does Not Allow to Draw Conclusions About Machine Lear ning Approaches,

    N. Reimers and I. Gurevych, “Why Comparing Single Perfo rmance Scores Does Not Allow to Draw Conclusions About Machine Lear ning Approaches,” arXiv preprint arXiv:1803.09578 , Mar 2018

    2018 arXiv

  28. [36]

    The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution,

    N. Nachar et al., “The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution,” Tutorials in quantitative Methods for Psychology , vol. 4, no. 1, pp. 13–20, Mar 2008

    2008

  29. [37]

    SOLAR 10.7B: Scaling Large Language Mod- els with Simple yet Effective Depth Up-Scaling,

    D. Kim et al. , “SOLAR 10.7B: Scaling Large Language Mod- els with Simple yet Effective Depth Up-Scaling,” arXiv preprint arXiv:2312.15166, Apr 2024

    2024 arXiv

  30. [38]

    Link Abstraction Models for Multicarrier Systems: A Logistic Regression Approach,

    C. Mesa et al. , “Link Abstraction Models for Multicarrier Systems: A Logistic Regression Approach,” International Journal of Communication Systems, vol. 31, no. 1, p. e3436, Sep 2017

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.