Prompting Wireless Networks: Reinforced In-Context Learning for Power Control

Chengming Hu; Di Wu; Dun Yuan; Hao Zhou; Jianzhong (Charlie) Zhang; Xue Liu; Ye Yuan

REVIEW 5 major objections 5 minor 24 references

Prompting Wireless Networks: Reinforced In-Context Learning for Power Control

T0 review · 5 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 This paper claims that prompting a large language model with a small set of past state-action-reward examples can match a trained deep Q-network on base-station power control, with no model training or fine-tuning.

desk verdict The idea is worth a look, but the paper doesn't yet show the LLM is doing the work—the nearest-example lookup is the real contender. read the letter →

arxiv 2506.06526 v1 pith:2LM5MNY4 submitted 2025-06-06 eess.SP

Hao Zhou , Chengming Hu , Dun Yuan , Ye Yuan , Di Wu , Xue Liu , Jianzhong (Charlie) Zhang This is my paper

classification eess.SP

keywords in-contextlearningpowercontrollargelanguagemodelswirelessnetworkoptimizationexampleselectionreinforcementbasestationpromptdesign

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

ProWin sets out to show that a large language model can solve a base-station power-control task from a natural-language description and a few curated past examples, with no model training or fine-tuning. The framework stores each past decision as a (state, action, reward) example, selects the most relevant examples for the current state, and inserts them into a written prompt that states the task goal, definition, and rules. In a three-small-cell case study, the paper reports that the most recent LLMs reach rewards and service quality comparable to a deep Q-network on both discrete user-number states and continuous average-distance states, while an older model trails. If the claim holds, network operators could optimize wireless resources by editing a prompt instead of training an agent or writing a new convex program, and get natural-language explanations of each decision.

What carries the argument

The load-bearing object is the reinforced in-context learning prompt: a natural-language task description combined with an example set chosen from an online experience pool. Selection is state-based for discrete states and ranking-based for continuous states, where the metric $L(E, s_{\mathrm{target}}) = r(s,a) - \tau \|s - s_{\mathrm{target}}\|$ balances past reward against state similarity, and an epsilon-greedy rule supplies exploration. The paper justifies the scheme with a linear-attention identity in which demonstration tokens contribute an implicit weight update $\Delta W_{ICL} = \sum_i (W_K E_i) \otimes (W_V E_i)$ to the zero-shot attention weights, making example selection the functional equivalent of a gradient step.

What would settle it

Run the Appendix B scenario at target distance 8.72 repeatedly and record every episode in which the model's explanation misorders distances (e.g., calling 7.76 farther than 8.72); if any such episode picks a power level other than the one the good examples imply, or yields a reward below the converged DQN value, the assumption that reasoning errors do not hurt decisions fails.

Watch

Extended reading notes

Core claim

The central claim, stated most strongly in the abstract, is that ProWin outperforms RL-based methods; the body reports the result in more measured terms as comparable performance to DQN. Each decision is made by prompting an LLM with a task description plus selected examples, and the resulting (state, action, reward) is appended to an experience pool for future prompts. Discrete states are handled by retrieving examples with the same state, continuous states by ranking examples with $L(E, s_{\mathrm{target}}) = r(s,a) - \tau \|s - s_{\mathrm{target}}\|$ and taking the top-$K$. The simulations show rewards and power consumption converging over episodes, behavior that adapts to stricter data-rate constraints, and performance that holds as the number of base stations grows.

Load-bearing premise

The claim rests on the LLM's chosen power level being correct even when its stated reasoning is wrong, because ProWin applies the selected level without any check against the justification or constraint.

Editorial extensions

If this is right

LLM inference can substitute for DRL training in a wireless power-control task, eliminating parameter updates and their compute cost.
New optimization tasks can be specified by editing the natural-language prompt rather than formulating a new convex program or training a new agent.
Keeping an experience pool and scoring candidate examples by reward minus state distance gives a practical recipe that improves with more examples until gains saturate.
The reported parity with DQN persists under stricter rate constraints and larger base-station counts, so the result is not tied to the exact simulated layout.

Reading between the lines

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

A natural extension the paper does not test is applying the same prompt-and-example loop to other continuous-state radio decisions, such as beamforming or bandwidth allocation, provided a scalar reward is available.
Because example selection already ranks past decisions by usefulness, a pure nearest-neighbor baseline without any LLM would isolate how much of the reported performance comes from the LLM's inference versus the selection rule itself.
The hallucination documented in Appendix B implies that the natural-language explanations should be treated as illustrative rather than auditable, unless a separate consistency check is added.
One testable corollary is that ProWin's advantage over older LLMs is mostly a model-capability effect; upgrading the LLM should improve decisions without changing any wireless-side code or example-selection rule.

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.

Desk Editor's Note

read the letter

Colleague—

Read this one with the stress-test note in hand, because the note is right. The headline claim that ProWin 'outperforms' DRL is contradicted by the body, which says 'comparable' in Sections 4.2 and 5. More importantly, the design has no ablation where the LLM is removed. Algorithm 1 already contains a complete non-LLM policy: pick the action of the example that maximizes r - tau*distance. For a 4-action problem, that lookup is likely near-optimal. The appendix even shows the LLM picking the right action for a wrong reason (7.76 misread as farther than 8.72), consistent with the examples, not the model's world model, driving the choice. So the central claim—that in-context learning replaces DRL—is unverified.

That said, the paper isn't empty. Applying reward-ranked in-context example selection to base-station power control is a genuine novelty, and the L(E, s_target) metric is a sensible way to trade reward and state similarity. The authors also deserve credit for documenting a hallucination and admitting it didn't hurt the final decision. That's honest. The no-finetuning angle is appealing for resource-constrained deployments.

Soft spots beyond the missing ablation: no error bars or multiple seeds; DQN architecture and hyperparameters are unreported; the 'reinforced' framing is oversold—there's no policy update, just epsilon-greedy exploration plus example selection on rewards, which is experience replay. The complexity analysis is hand-wavy about LLM inference cost, which dominates. And the abstract overstates the results: 'outperforms' should be 'matches.' These are fixable.

Who is this for? Readers working on LLM-for-wireless who want a clean demonstration that prompting with a few good examples can handle a small control problem. It's a workshop-level paper with a solid seed of an idea. It deserves a serious referee, but the referee should require a no-LLM ablation and a proper statistical treatment before publication. If that ablation shows the lookup alone matches the curves, then the paper is really about example selection, not ICL.

Referee Report

5 major / 5 minor

Summary. The paper proposes ProWin, a framework that uses in-context learning with large language models (LLMs) for base station power control, without model training or fine-tuning. The method constructs natural-language task prompts and augments them with examples drawn from an experience pool; for discrete states examples are selected by exact state matching (Section 3.3), while for continuous states they are chosen by the reward-distance metric of Eq. (11) (Section 3.4). An epsilon-greedy rule (Eq. (10)) mixes random exploration with LLM-based decisions, and each executed state-action-reward tuple is appended to the pool. The evaluation considers three LLMs (Llama3-8B, Llama3-70B, GPT-3.5 Turbo) on two scenarios (discrete user-number states and continuous average-distance states) against a deep Q-network baseline. The authors report comparable performance to DQN in the body and claim in the abstract that ProWin outperforms RL-based methods. An appendix documents a hallucination case where Llama3-8B misreads distances yet still selects the correct power level.

Significance. If the central claim were fully established, the paper would show that an off-the-shelf LLM, guided by a handful of rewarded examples, can match a trained deep RL agent on a wireless power-control task while also providing natural-language explanations. This would be a valuable result for resource-constrained network optimization, and the proposed framework is simple and easy to reproduce in spirit. The paper honestly includes a hallucination analysis in Appendix B and evaluates three different LLMs, which is a strength. However, the significance as stated is conditional: the experimental design does not isolate the contribution of the LLM from the example-selection mechanism, the abstract overstates the body's findings, and the missing variance information makes the 'comparable' claim difficult to assess. The core idea of reward-weighted example selection is interesting, but the evidence currently does not support the strong interpretation that LLM in-context reasoning is the effective decision-maker.

major comments (5)

[Abstract, Section 4.2, Section 5] The abstract states that ProWin 'outperforms reinforcement learning (RL)-based methods,' but Section 4.2 says that the Llama3 models 'achieve a comparable reward and service quality as the DRL baseline' and Section 5 concludes with 'comparable performance as conventional DRL algorithms.' This is a direct mismatch between the paper's headline claim and its reported results. Please either present statistically supported evidence of superiority (with error bars and significance tests) or revise the abstract to claim comparable performance.
[Section 3.4, Eq. (11), Algorithm 1] The paper does not include any ablation that isolates the role of the LLM. The metric L(E, starget) = r(s,a) - tau * ||s - starget|| defines a complete non-LLM policy: at a state starget, one can output the action of the example that maximizes L, and for discrete states one can output the highest-reward action among examples with s = starget. The prompt construction in Algorithm 1 already contains this decision rule, and with only four power levels it is plausible that this lookup baseline reproduces the reported reward curves. Without a no-LLM baseline, a measurement of how often the LLM deviates from the Eq. (11) argmax, or an analysis of which examples are selected, the central claim that in-context learning—rather than the example-selection mechanism—is responsible for the performance is not established. This is the most load-bearing gap in the paper.
[Section 4.1, Section 4.2, Figures 3–10] The experimental evaluation lacks variance information. All learning curves and bar charts appear to be single-run results with no error bars, no confidence intervals, and no statement about the number of independent seeds. The DQN baseline is described in only two sentences in Section 4.1, with no architecture, hyperparameters, replay buffer size, learning rate, or training length. Since the paper's quantitative claim is precisely that ProWin is comparable to or better than DQN, the absence of any measure of variability makes this comparison unverifiable. Please report multi-seed averages with standard deviations and give a complete DQN configuration.
[Section 4.1, Algorithm 1, Eqs. (8), (10), (11)] Several algorithm parameters that directly affect the results are never given numerical values or a sensitivity analysis: P_target and beta in Eq. (8), epsilon in Eq. (10), tau in Eq. (11), the number of examples K, and the constraint C_min. Algorithm 1 lists these as inputs, but Section 4.1 does not state how they were set. The performance of the example-selection rule is likely sensitive to tau and K, and the reward definition depends on P_target and beta. Without these values the experiments are not reproducible, and the robustness of the reported 'comparable' performance cannot be assessed.
[Appendix B, Section 1] A stated motivation of the work is interpretability and transparency of LLM-based decisions, but Appendix B documents a concrete case where Llama3-8B misreads the distances 7.76 and 8.12 as 'farther away' than 8.72, yet still selects the correct power level. The algorithm has no guardrail to detect or correct such episodes, and the paper does not quantify how often such hallucinations occur or whether they ever lead to wrong actions. This episode is also consistent with the hypothesis that the prompt examples, rather than the LLM's reasoning, drive the final action. The interpretability advantage is therefore not currently demonstrated, and the framework needs either a consistency check on the LLM's stated rationale or a clear statement that explanations may be unreliable.

minor comments (5)

[Section 3.1, Eqs. (4)–(6)] The linear-attention analysis of in-context learning closely follows the derivation in Dai et al. (2022) and is presented as a proof ('It proves that...'). This is a restatement of prior work rather than a new theoretical contribution; the authors should either cite the derivation as prior analysis or explain what new insight Eq. (6) provides for the proposed selection scheme.
[Section 2.1, Eq. (1)] Equation (1) has typesetting errors: the first summation appears as 'KbP k=1' and the interference term lacks clear index notation, making the formula difficult to parse. Please correct and clean up the equation.
[Appendix A] In the illustrative prompt, 'Example 49' appears twice with different average user distances (8.78 and 8.96). If this is not a typographical error, the duplicate indexing should be clarified.
[Sections 3.3 and 3.4] The term 'reinforced in-context learning' may mislead readers, since Algorithm 1 does not perform a policy update or gradient-based reinforcement learning; the method is experience replay with reward-weighted example selection. Consider using a less loaded term or clearly stating that 'reinforced' refers to the reward-guided selection of examples.
[Figures 6–9] The captions of Figures 6–9 do not define what each bar or line represents or whether error bars are present. For example, Figure 9 is described as a 'comparison' but appears to show only the proposed method. Please make the captions self-contained and clarify whether any baselines are included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the example selection and evaluation use the same reward function, but this is standard online learning rather than a derivation that reduces to its inputs.

full rationale

The paper's claimed contribution is an empirical framework (ProWin) that selects demonstration examples by reward and state similarity (Eq. 11) and feeds them to an LLM, whose output is then evaluated in closed-loop simulation. No step in the derivation chain makes a predicted quantity equal to an input by construction. The example-selection metric uses the same reward function r(s,a) as the evaluation metric, but this is normal for an optimization algorithm: the method is designed to maximize that reward, and the experiments measure the resulting reward, so there is no fitted parameter being relabeled as a prediction. The theoretical argument in Sec. 3.1 that examples affect in-context learning weights is imported from an external citation (Dai et al. 2022) and is not self-referential. Self-citations (Zhou et al. 2022 for the rate equation and the DQN baseline) are not load-bearing; the rate equation is standard and DQN is a standard baseline. Appendix B candidly documents an LLM hallucination, which weakens the interpretability and robustness claims but does not make the derivation circular. The absence of an ablation that replaces the LLM with a direct argmax of Eq. (11) is a legitimate experimental-design concern about whether the LLM is necessary, but it is not circularity of the kind defined here, because the paper never claims to derive the LLM's output from Eq. (11) by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-chosen but unreported hyperparameters (tau, beta, epsilon, K, P_target) and on domain simplifications (equal power allocation, OFDMA, stationary environment). No new physical or algorithmic entities are introduced.

free parameters (5)

tau (weighting factor in Eq. 11) = not reported
Balances reward against state distance in example ranking; the paper gives no value or sensitivity analysis.
beta (penalty term in Eq. 8) = not reported
Penalizes constraint violations in the reward; value affects which examples are labeled good or bad.
epsilon (epsilon-greedy probability) = not reported
Controls exploration vs exploitation; affects the content of the experience pool.
K (number of examples in prompt) = not reported
Varies in Figure 10 but no absolute values are given; directly determines prompt length and LLM input.
P_target (target power in Eq. 8) = not reported
Defines the reward scale and thus the ranking of examples.

assumptions (4)

standard math Linear attention is a valid approximation of real attention for the in-context learning analysis (Eq. 5)
Used only for qualitative explanation; the algorithm does not rely on this derivation.
domain assumption OFDMA eliminates intra-cell interference and power is equally allocated across resource blocks
Section 2.1 states these simplifications before formulating the objective.
domain assumption The environment is stationary enough that historical examples remain relevant to future states
The example pool assumes past state-action-reward tuples generalize to new states; no non-stationarity handling is provided.
domain assumption The LLM output can be parsed into one of the four power levels
The algorithm assumes valid responses; parsing errors or off-format replies are not handled.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Prompting Wireless Networks: Reinforced In-Context Learning for Power Control." pith.science (2026). https://pith.science/paper/2LM5MNY4

@misc{pith2026250606526,
  author       = {Pith},
  title        = {Pith review of: Prompting Wireless Networks: Reinforced In-Context Learning for Power Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LM5MNY4}},
  note         = {Machine review of arXiv:2506.06526}
}

read the original abstract

To manage and optimize constantly evolving wireless networks, existing machine learning (ML)- based studies operate as black-box models, leading to increased computational costs during training and a lack of transparency in decision-making, which limits their practical applicability in wireless networks. Motivated by recent advancements in large language model (LLM)-enabled wireless networks, this paper proposes ProWin, a novel framework that leverages reinforced in-context learning to design task-specific demonstration Prompts for Wireless Network optimization, relying on the inference capabilities of LLMs without the need for dedicated model training or finetuning. The task-specific prompts are designed to incorporate natural language descriptions of the task description and formulation, enhancing interpretability and eliminating the need for specialized expertise in network optimization. We further propose a reinforced in-context learning scheme that incorporates a set of advisable examples into task-specific prompts, wherein informative examples capturing historical environment states and decisions are adaptively selected to guide current decision-making. Evaluations on a case study of base station power control showcases that the proposed ProWin outperforms reinforcement learning (RL)-based methods, highlighting the potential for next-generation future wireless network optimization.

Figures

Figures reproduced from arXiv: 2506.06526 by the authors.

**Figure 1.** Overall design of the proposed LLM-enabled in-context learning for transmission power control. We define WZSL = (WKQ) T WV Q for ease of notations, since the WZSL parameters include the query token Q only. Therefore, WZSL indicates the zero-shot learning case without examples. Then we can rewrite equation (5) as ˜fICL(q) = qWZSL + q (WKE) T WV E = qWZSL + LinearAttention (q, WKE, WV E) = qWZSL + q X i (WKEi) T WV… view at source ↗

**Figure 2.** The overall procedure of the example-related scheme. rithm 1. With the epsilon-greedy policy, the algorithm can try different actions by random exploration, finding better network optimization decisions. On the other hand, LLM can utilize the accumulated experience and examples to make good decisions, balancing exploration and exploitation. Such a design indicates that LLMs can constantly explore the network envir… view at source ↗

**Figure 3.** [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Continuous state space: System reward comparison of various LLMs. instruct, Llama3-70b-instruct, and GPT-3.5 turbo. Llama3- 8b is a small-scale LLM, while Llama3-70b and GPT-3.5 turbo are large models. Using LLM models with various sizes can better evaluate the capabil…

**Figure 5.** Figure 5: Continuous state space: Power consumption comparison of various LLMs. 4.2. Simulation Results [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 7.** Figure 7: Continuous state space: Average power consumption comparison under different data rate constraints. lower service quality, and higher power consumption. They demonstrate that the proposed in-context learning can adapt to different optimization settings and then adjust …

**Figure 10.** Figure 10: Reward performance with increasing number of examples and larger spaces. network performance iteratively, and adapt to different network environments. It achieves comparable performance as existing benchmarks DQN, and avoids the complexity of dedicated model paramet…

**Figure 9.** Figure 9: Continuous state space: Average service quality comparison with increasing number of BSs. 5) Effect of the number of examples: In [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 11 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Understanding telecom language through large language models

Bariah, L., Zou, H., Zhao, Q., Mouhouche, B., Bader, F., and Debbah, M. Understanding telecom language through large language models. In GLOBECOM 2023-2023 IEEE Global Communications Conference, pp.\ 6542--6547. IEEE, 2023

work page 2023
[3]

and Huber, M

Burkart, N. and Huber, M. F. A survey on the explainability of supervised machine learning. Journal of Artificial Intelligence Research, 70: 0 245--317, 2021

work page 2021
[4]

How many demonstrations do you need for in-context learning? arXiv preprint arXiv:2303.08119, 2023

Chen, J., Chen, L., Zhu, C., and Zhou, T. How many demonstrations do you need for in-context learning? arXiv preprint arXiv:2303.08119, 2023

arXiv 2023
[5]

H., Chong, K

Chen, Z., Yang, H. H., Chong, K. F. E., and Quek, T. Q. Personalizing semantic communication: A foundation model approach. In 2024 IEEE 25th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pp.\ 846--850. IEEE, 2024

work page 2024
[6]

W., et al

Chiang, M., Hande, P., Lan, T., Tan, C. W., et al. Power control in wireless cellular networks. Foundations and Trends in Networking , 2 0 (4): 0 381--533, 2008

work page 2008
[7]

Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers

Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., and Wei, F. Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. arXiv preprint arXiv:2212.10559, 2022

arXiv 2022
[8]

A survey on in-context learning

Dong, Q., Li, L., Dai, D., Zheng, C., Wu, Z., and et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022

arXiv 2022

Show all 24 references

[9]

Pushing large language models to the 6G edge: Vision, challenges, and opportunities

Lin, Z., Qu, G., Chen, Q., Chen, X., Chen, Z., and Huang, K. Pushing large language models to the 6G edge: Vision, challenges, and opportunities. arXiv preprint arXiv:2309.16739, 2023

2023 arXiv
[10]

A., and Yu, W

Liu, Y.-F., Chang, T.-H., Hong, M., Wu, Z., Man-Cho So, A., Jorswieck, E. A., and Yu, W. A survey of recent advances in optimization methods for wireless communications. IEEE Journal on Selected Areas in Communications, 42 0 (11): 0 2992--3031, 2024. doi:10.1109/JSAC.2024.3443759

2024
[11]

Large language models for telecom: Forthcoming impact on the industry

Maatouk, A., Piovesan, N., Ayed, F., De Domenico, A., and Debbah, M. Large language models for telecom: Forthcoming impact on the industry. IEEE Communications Magazine, 2024

2024
[12]

Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837, 2022

Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., and Zettlemoyer, L. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837, 2022

2022 arXiv
[13]

Q., and Kim, S.-L

Oh, S., Kim, J., Park, J., Ko, S.-W., Quek, T. Q., and Kim, S.-L. Uncertainty-aware hybrid inference with on-device small and remote large language models. arXiv preprint arXiv:2412.12687, 2024

2024 arXiv
[14]

Training language models to follow instructions with human feedback

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

2022
[15]

Towards semantic mac protocols for 6g: From protocol learning to language-oriented approaches

Park, J., Ko, S.-W., Choi, J., Kim, S.-L., Choi, J., and Bennis, M. Towards semantic mac protocols for 6g: From protocol learning to language-oriented approaches. IEEE BITS the Information Theory Magazine, 2024

2024
[16]

Large language model-based wireless network design

Qiu, K., Bakirtzis, S., Wassell, I., Song, H., Zhang, J., and Wang, K. Large language model-based wireless network design. IEEE Wireless Communications Letters, 2024

2024
[17]

Federated llms fine-tuned with adaptive importance-aware lora

Su, Y., Yan, N., and Deng, Y. Federated llms fine-tuned with adaptive importance-aware lora. arXiv preprint arXiv:2411.06581, 2024

2024 arXiv
[18]

The learnability of in-context learning

Wies, N., Levine, Y., and Shashua, A. The learnability of in-context learning. Advances in Neural Information Processing Systems, 36: 0 36637--36651, 2023

2023
[19]

and Salim, F

Xue, H. and Salim, F. D. Promptcast: A new prompt-based learning paradigm for time series forecasting. IEEE Transactions on Knowledge and Data Engineering, pp.\ 1--14, 2023. doi:10.1109/TKDE.2023.3342137

2023
[20]

Federated fine-tuning of llms: Framework comparison and research directions

Yan, N., Su, Y., Deng, Y., and Schober, R. Federated fine-tuning of llms: Framework comparison and research directions. arXiv preprint arXiv:2501.04436, 2025

2025 arXiv
[21]

and Liang, Y.-C

Zhang, L. and Liang, Y.-C. Deep reinforcement learning for multi-agent power control in heterogeneous networks. IEEE Transactions on Wireless Communications, 20 0 (4): 0 2551--2564, 2020

2020
[22]

K., and Fan, P

Zhang, Z., Xiao, Y., Ma, Z., Xiao, M., Ding, Z., Lei, X., Karagiannidis, G. K., and Fan, P. 6g wireless networks: Vision, requirements, architecture, and key technologies. IEEE vehicular technology magazine, 14 0 (3): 0 28--41
[23]

Zhou, H., Erol-Kantarci, M., and Poor, H. V. Learning from peers: Deep transfer reinforcement learning for joint radio and cache resource allocation in 5G RAN slicing. IEEE Transactions on Cognitive Communications and Networking, 8 0 (4): 0 1925--1941, 2022

1925
[24]

Zhou, H., Erol-Kantarci, M., Liu, Y., and Poor, H. V. A survey on model-based, heuristic, and machine learning optimization approaches in ris-aided wireless networks. IEEE Communications Surveys & Tutorials, 2023

2023

Pith tools

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

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

Understanding telecom language through large language models

Bariah, L., Zou, H., Zhao, Q., Mouhouche, B., Bader, F., and Debbah, M. Understanding telecom language through large language models. In GLOBECOM 2023-2023 IEEE Global Communications Conference, pp.\ 6542--6547. IEEE, 2023

work page 2023

[3] [3]

and Huber, M

Burkart, N. and Huber, M. F. A survey on the explainability of supervised machine learning. Journal of Artificial Intelligence Research, 70: 0 245--317, 2021

work page 2021

[4] [4]

How many demonstrations do you need for in-context learning? arXiv preprint arXiv:2303.08119, 2023

Chen, J., Chen, L., Zhu, C., and Zhou, T. How many demonstrations do you need for in-context learning? arXiv preprint arXiv:2303.08119, 2023

arXiv 2023

[5] [5]

H., Chong, K

Chen, Z., Yang, H. H., Chong, K. F. E., and Quek, T. Q. Personalizing semantic communication: A foundation model approach. In 2024 IEEE 25th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pp.\ 846--850. IEEE, 2024

work page 2024

[6] [6]

W., et al

Chiang, M., Hande, P., Lan, T., Tan, C. W., et al. Power control in wireless cellular networks. Foundations and Trends in Networking , 2 0 (4): 0 381--533, 2008

work page 2008

[7] [7]

Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers

Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., and Wei, F. Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. arXiv preprint arXiv:2212.10559, 2022

arXiv 2022

[8] [8]

A survey on in-context learning

Dong, Q., Li, L., Dai, D., Zheng, C., Wu, Z., and et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022

arXiv 2022