Chain-of-Thought for Large Language Model-empowered Wireless Communications

Dusit Niyato; Hongyang Du; Jiacheng Wang; Jian Zhu; Lei Feng; Ruichen Zhang; Shiwen Mao; Xudong Wang; Zhu Han

REVIEW 4 major objections 6 minor 5 cited by

Chain-of-Thought for Large Language Model-empowered Wireless Communications

T0 review · 4 major / 6 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 prompting a large language model to reason step by step, rather than answer directly, yields better wireless control policies, with about a 27.2% higher sum rate in a UAV deployment and power-allocation case study.

desk verdict The CoT-vs-baseline claim is under-evidenced (unspecified baseline prompt, n=10, no variance), but the survey and framework are useful for readers entering the area. read the letter →

arxiv 2505.22320 v1 pith:O2F2PZYO submitted 2025-05-28 cs.NI

Xudong Wang , Jian Zhu , Ruichen Zhang , Lei Feng , Dusit Niyato , Jiacheng Wang , Hongyang Du , Shiwen Mao

show 1 more author

Zhu Han

This is my paper · ORCID

classification cs.NI

keywords chain-of-thoughtpromptinglargelanguagemodelsintent-drivennetworkingwirelesscommunicationsUAVnetworksresourceallocationreinforcementlearningAuto-CoT

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

Chain-of-thought (CoT) prompting is usually studied on language and math puzzles; this paper argues it also changes how well a large language model can run a wireless network. The authors review CoT variants, then build a three-layer intent-driven framework that turns a user's natural-language goal into a concrete control policy: intents are embedded and clustered, a reinforcement-learning agent picks a CoT reasoning module, the LLM reasons step by step, and the resulting trace is parsed into wireless parameters. In a UAV base-station case study, the CoT configuration using GPT-4o reaches about a 27.2% higher sum rate than a no-CoT baseline at 400 m communication range, with higher coverage and better reasoning-quality scores. If that comparison is clean, the practical consequence is that part of wireless network optimization can be shifted from retraining models to structuring their prompts.

What carries the argument

The load-bearing object is the CoT-enabled decision layer with its modular Auto-CoT prompting pipeline. User intents are embedded with Sentence-BERT and clustered with K-Means; a deep Q-network selects the most relevant CoT module for the request; each module contains hand-authored few-shot exemplars that force a fixed reasoning sequence (interpret intent, extract coordinates and channel parameters, formulate the FSPL/SINR/Shannon optimization, solve heuristically, compute utility); and a semantic parser converts the reasoning trace into executable transmit-power and UAV-placement commands. The loop is scored by a composite utility $Q_{\text{total}} = \alpha Q_{\text{LLM}} + \beta(Q_c + Q_R)$, which is what the DRL agent is trained to maximize, so the reasoning quality and the wireless performance are optimized jointly.

What would settle it

Run the UAV case study with three GPT-4o prompts: the full CoT prompt, a control prompt containing the same worked examples and constraints but instructing the model to output only the final deployment and power values, and a bare prompt with no examples. If the exemplar-only control matches or beats the full CoT prompt, the central claim fails; if the full CoT prompt wins, the claim survives. The paper's own Fig. 4 data provide the baseline numbers to compare against.

Watch

Extended reading notes

Core claim

The paper's central claim is that explicit intermediate reasoning steps are not a cosmetic addition to LLM-based wireless control; they are the mechanism that produces better decisions. Concretely, the authors report that the intent-driven framework with Auto-CoT reasoning outperforms the same framework without CoT across UAV communication ranges from 200 m to 550 m, and that at 400 m the sum rate improves by approximately 27.2%. They also report higher system utility, coverage rate, and LLM fitness scores, and that pairing a stronger LLM (GPT-4o over GPT-3.5) and a DRL-based module selector over a random selector improves every measured metric. The intended upshot is that CoT gives wireless operators both interpretability and performance: the model explains its reasoning while also producing policies that score better on physical-layer metrics.

Load-bearing premise

The case study assumes the comparison isolates CoT: the no-CoT baseline must be identical to the CoT version except for the missing intermediate reasoning steps, but the paper never specifies the baseline prompt, so the 27.2% gain could come from the extra worked examples rather than from reasoning per se.

Editorial extensions

If this is right

If the case study generalizes, prompt-level CoT can improve deployed LLM wireless controllers without fine-tuning or retraining.
Natural-language intents can be mapped to concrete wireless parameters, so operators can steer networks with high-level goals rather than numeric configuration files.
The DRL-based module selector result implies that choosing which reasoning module to use is itself a learnable policy, not a fixed design choice.
The reported gains at every tested UAV communication range imply CoT helps across coverage regimes, not only at one operating point.

Reading between the lines

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

The paper's implicit promise is that the 27.2% gain is attributable to step-by-step reasoning; a natural extension would test a control prompt containing the same worked examples but asking for a direct final answer, which would separate CoT from exemplar content.
If the mechanism is real, CoT-style wrappers should transfer to other LLM-based wireless tasks such as network slicing, handover, and spectrum sharing, where the paper's survey predicts similar gains but provides no measurements.
The composite utility couples reasoning quality to wireless performance, which suggests a testable design space: changing the weights $\alpha$ and $\beta$ should trade off interpretability against physical-layer gain in a predictable way.
RL-driven module selection over a library of reasoning chains hints that future systems could evolve their own reasoning templates, beyond the hand-authored exemplars in Fig. 3; the paper does not test that, but its architecture makes it possible.

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 — you should know this paper up front: the survey half is fine, but the central empirical claim — that CoT gives roughly a 27% sum-rate gain over a non-CoT baseline — isn't yet supported by the evidence. The baseline's prompt is never specified beyond "GPT-4o without CoT," so the comparison may be prompt-exemplar engineering rather than CoT reasoning. And results are averaged over 10 random layouts with no error bars or significance tests.

What's actually new: a three-layer intent-driven CoT framework that uses Sentence-BERT clustering, DRL-based module activation, and Auto-CoT, applied to a UAV deployment and power-control case study. The framework is a sensible combination of known techniques; the survey of CoT types and wireless applications is organized and readable, and Table I is a decent map of the area. The case study is genuinely new, but it reads as a demonstration, not an evaluation.

Where it's soft, in proportion: the missing baseline specification is the biggest issue. Without the exact non-CoT prompt, the 27.2% gain isn't attributable to CoT. Second, no variance reporting — with n=10, the "significantly outperforms" language is unsupported. Third, the composite utility Qtotal includes QLLM, a fitness score that rewards informativeness and consistency; CoT tends to produce longer, more structured output, so part of the "reasoning quality" gain is baked into the metric. The coverage and sum-rate comparisons are less circular, but they still depend on the uncontrolled prompt comparison. There's also a small arithmetic inconsistency in the worked example in Fig. 1 — minor, but sloppy. No code or prompts are released, which makes the whole thing hard to reproduce.

If the authors add a controlled ablation (same exemplars, no intermediate reasoning steps), report variance over layouts, and release prompts and code, the case study could support a modest claim. As written, it's a useful framework plus a pilot result, not a demonstrated one.

Who gets value: readers wanting a compact survey of CoT variants and LLM+wireless applications, or people building intent-driven frameworks who will run their own controlled experiments. I'd send this to peer review — a serious referee can force the needed ablations — but I wouldn't cite the empirical result as it stands.

Referee Report

4 major / 6 minor

Summary. The paper reviews Chain-of-Thought (CoT) techniques and their applications in wireless communications, then proposes a multi-layer intent-driven CoT framework that parses user intents, clusters them, activates reasoning modules via deep reinforcement learning, generates interpretable control policies, and evaluates them with a composite utility. A UAV deployment and power-control case study with GPT-4o and GPT-3.5 reports that the CoT-enabled framework outperforms a non-CoT baseline, including a roughly 27.2% sum-rate gain at 400 m, alongside improved coverage and reasoning quality.

Significance. If the reported gains hold, the paper would provide useful evidence that CoT prompting can improve LLM-generated wireless control policies and would contribute a reusable framework for intent-driven networks. The survey portion is well organized and the case study makes a falsifiable, quantitative claim. However, the empirical support is currently too thin to carry the headline claim: the non-CoT baseline is not specified, no variance or significance information is reported for the 10 random layouts, and the reasoning-quality metric may reward CoT by construction. The authors should be credited for a clearly presented framework and for making the central comparison explicit, but the evidence needs substantial strengthening.

major comments (4)

[Sec. IV-C, Fig. 4] The non-CoT baseline is not defined. The text only says "GPT-4o without CoT" and states that both configurations use DRL-enhanced module activation; it does not specify the baseline prompt, whether the baseline also receives the hand-authored exemplars shown in Fig. 3, or how parsing and post-processing differ. Because the CoT prompt contains a fully worked numerical example (Q1/A1), the reported 27.2% gain could be due to exemplar quality or prompt engineering rather than to intermediate reasoning. A controlled ablation that keeps exemplars, module activation, and parsing identical while removing only the step-by-step reasoning trace is necessary to support the central claim.
[Sec. IV-B, Sec. IV-C] All reported values are averages over 10 random user-location samples, but no standard deviations, confidence intervals, or significance tests are given anywhere. With n = 10, a 27.2% difference in Fig. 4 may be within sampling noise, so the word "significantly" in the abstract and in Section IV-C is not yet supported. Please report the per-sample results, add error bars, and provide a paired significance test across the same 10 layouts.
[Step 5, Sec. IV-B] The composite utility Qtotal = αQLLM + β(Qc + QR) includes a fitness score QLLM that rewards informativeness and consistency. Since CoT is designed to produce longer, more structured outputs, part of the reported improvement in reasoning quality is built into the metric. Moreover, Fig. 4 reports no QLLM comparison for the non-CoT baseline, so the claim of improved "quality of generated reasoning" is not empirically demonstrated. The authors should either use a metric that does not trivially favor verbose/structured text or report the components of QLLM separately for both conditions.
[Sec. IV-B, Fig. 5] The DRL-based module activation is described only at a high level: the state space, action space, reward definition (beyond the composite utility), training procedure, and convergence behavior are not specified. This matters because Fig. 5 attributes performance differences to DRL activation versus random activation, so the reader cannot determine whether the comparison isolates CoT, model choice, or DRL training. Please provide the missing implementation details or explicitly state that the DRL component is an illustrative module whose individual contribution is not separately evaluated.

minor comments (6)

[Fig. 1] In the worked example, the noise power is written as "n=1 dB" and the displayed rates are consistent only if n is interpreted in linear units (10^{0.1} ≈ 1.2589); please state the units explicitly. In addition, the phrase "power allocation based on the channel gain" is ambiguous: the displayed p' is obtained by subtracting gain-proportional reductions from the initial powers, not by allocating the 14 W total proportionally to the gains; if the latter was intended, p' should be (4.6667, 7.4667, 1.8667).
[Sec. IV-B] The exact formulas used to compute SINR, Shannon capacity, and noise power are not given; please include them so that the case study is reproducible.
[Sec. I, Fig. 1] There are several typos: "This fundamentals offer" should be "These fundamentals offer"; "UA V" appears with inconsistent spacing in multiple places; and "Automatic reason chain generation" in Fig. 1 should be "Automatic reasoning chain generation."
[Fig. 2 caption] The caption directs readers to an external GitHub repository for references [A1-A11] and [B1-B4]; these references should be included in the main reference list rather than only appearing in an external link.
[Step 5, Sec. IV-B] The normalization of Qc, QR, and QLLM is not defined, and the relationship between Qtotal = αQLLM + βQwireless in Step 5 and Qtotal = αQLLM + β(Qc + QR) in Section IV-B is not explained; please clarify how the components are normalized and combined.
[Fig. 5] The three configurations in Fig. 5 are all CoT-enhanced, so this figure does not provide evidence about the non-CoT comparison; consider adding the non-CoT condition to Fig. 5 or explaining how Fig. 4 and Fig. 5 jointly support the claims.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central 27.2% sum-rate gain is an empirical LLM prompt comparison, not a quantity derived from its own definitions; the underspecified non-CoT baseline is an experimental-control concern, not a circularity.

full rationale

The paper's load-bearing numerical claim is an observed difference between GPT-4o with and without CoT prompting in a UAV network case study, not a first-principles derivation. The framework's composite utility Qtotal includes a QLLM fitness score from [15] that rewards consistency and informativeness; however, the headline Fig. 4 comparison reports coverage and sum rate, which do not depend on QLLM, and Fig. 5's fitness-score comparison is among CoT variants only. Therefore any alignment between QLLM and CoT-style outputs does not, by construction, produce the reported 27.2% sum-rate gain. The self-citations ([5], [13], [14]) appear in survey or motivational contexts and are not load-bearing for the case-study conclusion. The main validity threats are missing details about the non-CoT prompt, lack of variance or significance reporting over n=10 random placements, and possible conflation of exemplar quality with CoT; these are methodological gaps rather than circular reductions. No equation in the paper is equivalent to its input by construction, and no fitted parameter is renamed as a prediction. Score 2 reflects only the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The central claim rests on hand-set utility weights, hand-authored prompts, an idealized channel model, and unvalidated framework components such as the DRL agent and semantic parsing engine. There is no calibration against external benchmarks. The only standard math inputs are free-space path loss and Shannon capacity, which are treated as given.

free parameters (3)

Utility weights α and β = α = 0.1, β = 0.45
Set by hand in Section IV-B; define the composite utility Qtotal = αQLLM + β(Qc + QR) used to evaluate and optimize strategies. Changing them changes which strategy is preferred.
CoT exemplar content = Hand-authored (e.g., warehouse AP power example in Fig. 3)
Few-shot exemplars are manually constructed, not systematically varied; they may encode the solution structure and drive the reported gains.
K-Means cluster count K = Not reported
Required by Step 1 for intent clustering; the case study does not report K, so the full framework is not evaluated end-to-end.

assumptions (6)

domain assumption Free-space path loss with no fading or shadowing adequately models UAV-to-user links
Entered in Section IV-B experimental setup; real UAV networks have multipath and obstruction.
domain assumption Uniform spatial distribution of users represents the deployment scenario
Section IV-B: users randomly distributed by uniform process; real user locations are often clustered.
domain assumption LLM-generated strategies can be parsed into executable wireless commands with acceptable fidelity
Step 4 claims a neural semantic parsing engine, but no implementation or error rate is given.
domain assumption The DQN agent can learn useful CoT module selection from the composite utility reward
Step 2 relies on DRL trained with Qtotal, but no training details or convergence results are provided.
domain assumption The fitness score from [15] is an unbiased measure of reasoning quality
Step 5 uses consistency, informativeness, and misleadingness; informativeness may favor lengthier CoT outputs.
standard math Shannon capacity with FSPL-based SINR is the correct rate model
Used throughout CoT reasoning and evaluation; standard in simplified link analysis.

invented entities (1)

Multi-layer intent-driven CoT framework
purpose: Translate natural-language user intents into executable wireless control policies via intent parsing, DRL module activation, CoT reasoning, and strategy extraction
Proposed architecture in Section IV-A; not implemented end-to-end and has no falsifiable handle outside the paper beyond a small simulation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Chain-of-Thought for Large Language Model-empowered Wireless Communications." pith.science (2026). https://pith.science/paper/O2F2PZYO

@misc{pith2026250522320,
  author       = {Pith},
  title        = {Pith review of: Chain-of-Thought for Large Language Model-empowered Wireless Communications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O2F2PZYO}},
  note         = {Machine review of arXiv:2505.22320}
}

read the original abstract

Recent advances in large language models (LLMs) have opened new possibilities for automated reasoning and decision-making in wireless networks. However, applying LLMs to wireless communications presents challenges such as limited capability in handling complex logic, generalization, and reasoning. Chain-of-Thought (CoT) prompting, which guides LLMs to generate explicit intermediate reasoning steps, has been shown to significantly improve LLM performance on complex tasks. Inspired by this, this paper explores the application potential of CoT-enhanced LLMs in wireless communications. Specifically, we first review the fundamental theory of CoT and summarize various types of CoT. We then survey key CoT and LLM techniques relevant to wireless communication and networking. Moreover, we introduce a multi-layer intent-driven CoT framework that bridges high-level user intent expressed in natural language with concrete wireless control actions. Our proposed framework sequentially parses and clusters intent, selects appropriate CoT reasoning modules via reinforcement learning, then generates interpretable control policies for system configuration. Using the unmanned aerial vehicle (UAV) network as a case study, we demonstrate that the proposed framework significantly outperforms a non-CoT baseline in both communication performance and quality of generated reasoning.

Figures

Figures reproduced from arXiv: 2505.22320 by the authors.

**Figure 1.** An example to illustrate the different chain-of-thought techniques. Zero-shot CoT generates reasoning chains without examples. Few-shot CoT provides exemplars to guide [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** The principles, advantages, disadvantages, and suitable applications of different chain-of-thought strategies, including Zero-shot CoT, Few-shot CoT, Auto-CoT, Self-consistency [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The proposed CoT-enabled IDNs framework operates across three layers: the application layer, the Chain-of-Thought-enabled decision layer, and the infrastructure layer. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Comprehensive performance comparison of the proposed CoT-driven framework [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Agentic AI for ISAC: Analysis, Framework, and Case Study
cs.AI 2025-12 reject novelty 5.0 of 10

An agentic ISAC framework using a transformer-based MoE policy and an LLM-designed reward reports 131% higher communication rate and 5.4% lower CRB than a SAC baseline in a small beamforming case study.
AI Reasoning for Wireless Communications and Networking: A Survey and Perspectives
cs.NI 2025-09 conditional novelty 4.0 of 10

A survey that organizes LLM and AI reasoning methods into a taxonomy and maps them onto the physical, link, network, transport, and application layers of wireless networks.
Large Language Models for Next-Generation Wireless Network Management: A Survey and Tutorial
cs.NI 2025-09 conditional novelty 4.0 of 10

A survey and tutorial that organizes LLM-enabled wireless network optimization into formulation, solution, and verification stages, with case studies drawn from the authors' own prior papers.
Toward Edge General Intelligence with Agentic AI and Agentification: Concepts, Technologies, and Future Directions
cs.NI 2025-08 conditional novelty 4.0 of 10

A survey that organizes agentic AI for 6G edge networks into four pillars, compactness, efficiency, knowledge and reasoning, and migration, and illustrates them with prior case studies.
Toward Edge General Intelligence with Multiple-Large Language Model (Multi-LLM): Architecture, Trust, and Orchestration
cs.NI 2025-07 conditional novelty 4.0 of 10

A survey of multi-LLM systems in edge computing, covering architectures, enabling technologies, trust mechanisms, applications, and open datasets for edge general intelligence.

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages · cited by 5 Pith papers

[1]

Large Language Model-Based Wireless Network Design,

K. Qiu, S. Bakirtzis, I. Wassell, H. Song, J. Zhang, and K. Wang, “Large Language Model-Based Wireless Network Design,”IEEE Wireless Com- munications Letters, vol. 13, no. 12, pp. 3340–3344, Dec. 2024. 0.6 0.4 900 0.7 0.3 7000.2 0.2 GPT-4o-DRL System utility (unitless) Fitness score (unitless) Coverage rate (%) Sum rate (Mbps)881.9 853.9812.0 0.617 0.556 ...

work page 2024
[2]

Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities,

H. Zhou, C. Hu, Y . Yuan, Y . Cui, Y . Jin, C. Chen, H. Wu, D. Yuan, L. Jiang, D. Wu, X. Liu, C. Zhang, X. Wang, and J. Liu, “Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities,” IEEE Com- munications Surveys & Tutorials , pp. 1–1, 2024

work page 2024
[3]

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,

J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” Advances in Neural Information Processing Systems, vol. 35, pp. 24 824–24 837, Dec. 2022

work page 2022
[4]

Graph of Thoughts: Solving Elaborate Problems with Large Language Models,

M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, M. Podstawski, L. Gianinazzi, J. Gajda, T. Lehmann, H. Niewiadomski, P. Nyczyk et al. , “Graph of Thoughts: Solving Elaborate Problems with Large Language Models,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 16, pp. 17 682–17 690, Vancouver, Canada, Feb. 2024

work page 2024
[5]

Wireless Hallucination in Generative AI-enabled Communications: Concepts, Issues, and Solutions,

X. Wang, J. Wang, L. Feng, D. Niyato, R. Zhang, J. Kang, Z. Xiong, H. Du, and S. Mao, “Wireless Hallucination in Generative AI-enabled Communications: Concepts, Issues, and Solutions,” arXiv preprint arXiv:2503.06149, 2025

arXiv 2025
[6]

Chain of Preference Optimization: Improving Chain-of-Thought Reasoning in LLMs,

X. Zhang, C. Du, T. Pang, Q. Liu, W. Gao, and M. Lin, “Chain of Preference Optimization: Improving Chain-of-Thought Reasoning in LLMs,” Advances in Neural Information Processing Systems , vol. 37, pp. 333–356, Dec. 2024

work page 2024
[7]

Large Language Models Are Zero-Shot Reasoners,

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large Language Models Are Zero-Shot Reasoners,” Advances in Neural In- formation Processing Systems , vol. 35, pp. 22 199–22 213, Dec. 2022

work page 2022
[8]

Automatic chain of thought prompting in large language models,

Z. Zhang, A. Zhang, M. Li, and A. Smola, “Automatic chain of thought prompting in large language models,”The Eleventh International Conference on Learning Representations , Kigali, Rwanda, May 2023

work page 2023

Show all 15 references

[9]

Self-Consistency Improves Chain of Thought Reasoning in Language Models,

X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-Consistency Improves Chain of Thought Reasoning in Language Models,” The Eleventh International Conference on Learning Representations , Kigali, Rwanda, May 2023

2023
[10]

Tree of Thoughts: Deliberate Problem Solving with Large Language Models,

S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan, “Tree of Thoughts: Deliberate Problem Solving with Large Language Models,” Advances in Neural Information Processing Systems, vol. 36, pp. 11 809–11 822, New Orleans, LA, USA, Dec. 2023

2023
[11]

SCoTT: Wireless-Aware Path Planning with Vision Language Models and Strate- gic Chains-of-Thought,

A. Djuhera, V . C. Andrei, A. Seffo, H. Boche, and W. Saad, “SCoTT: Wireless-Aware Path Planning with Vision Language Models and Strate- gic Chains-of-Thought,” arXiv preprint arXiv:2411.18212 , 2024

2024 arXiv
[12]

Large Language Models for Wireless Networks: An Overview from the Prompt Engineering Perspective,

H. Zhou, C. Hu, D. Yuan, Y . Yuan, D. Wu, X. Chen, H. Tabassum, and X. Liu, “Large Language Models for Wireless Networks: An Overview from the Prompt Engineering Perspective,” IEEE Wireless Communications, pp. 1–9, 2025

2025
[13]

Rethinking Generative Semantic Communica- tion for Multi-user Systems with Multi-Modal LLM,

W. Yang, Z. Xiong, S. Mao, T. Q. Quek, P. Zhang, M. Deb- bah, and R. Tafazolli, “Rethinking Generative Semantic Communica- tion for Multi-user Systems with Multi-Modal LLM,” arXiv preprint arXiv:2408.08765, 2024

2024 arXiv
[14]

Large Language Models Empower Multimodal Integrated Sensing and Communication,

L. Cheng, H. Zhang, B. Di, D. Niyato, and L. Song, “Large Language Models Empower Multimodal Integrated Sensing and Communication,” IEEE Communications Magazine , vol. 63, no. 5, pp. 190–197, May 2025

2025
[15]

Evolving Chain-of-Thought And Automatic Text Annotation Based Intent Classification Method for Wireless Network,

B. Li, T. Liu, J. Ma, J. Dai, and P. Liao, “Evolving Chain-of-Thought And Automatic Text Annotation Based Intent Classification Method for Wireless Network,” Computer Networks, vol. 259, p. 111074, 2025

2025

Pith tools

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

[1] [1]

Large Language Model-Based Wireless Network Design,

K. Qiu, S. Bakirtzis, I. Wassell, H. Song, J. Zhang, and K. Wang, “Large Language Model-Based Wireless Network Design,”IEEE Wireless Com- munications Letters, vol. 13, no. 12, pp. 3340–3344, Dec. 2024. 0.6 0.4 900 0.7 0.3 7000.2 0.2 GPT-4o-DRL System utility (unitless) Fitness score (unitless) Coverage rate (%) Sum rate (Mbps)881.9 853.9812.0 0.617 0.556 ...

work page 2024

[2] [2]

Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities,

H. Zhou, C. Hu, Y . Yuan, Y . Cui, Y . Jin, C. Chen, H. Wu, D. Yuan, L. Jiang, D. Wu, X. Liu, C. Zhang, X. Wang, and J. Liu, “Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities,” IEEE Com- munications Surveys & Tutorials , pp. 1–1, 2024

work page 2024

[3] [3]

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,

J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” Advances in Neural Information Processing Systems, vol. 35, pp. 24 824–24 837, Dec. 2022

work page 2022

[4] [4]

Graph of Thoughts: Solving Elaborate Problems with Large Language Models,

M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, M. Podstawski, L. Gianinazzi, J. Gajda, T. Lehmann, H. Niewiadomski, P. Nyczyk et al. , “Graph of Thoughts: Solving Elaborate Problems with Large Language Models,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 16, pp. 17 682–17 690, Vancouver, Canada, Feb. 2024

work page 2024

[5] [5]

Wireless Hallucination in Generative AI-enabled Communications: Concepts, Issues, and Solutions,

X. Wang, J. Wang, L. Feng, D. Niyato, R. Zhang, J. Kang, Z. Xiong, H. Du, and S. Mao, “Wireless Hallucination in Generative AI-enabled Communications: Concepts, Issues, and Solutions,” arXiv preprint arXiv:2503.06149, 2025

arXiv 2025

[6] [6]

Chain of Preference Optimization: Improving Chain-of-Thought Reasoning in LLMs,

X. Zhang, C. Du, T. Pang, Q. Liu, W. Gao, and M. Lin, “Chain of Preference Optimization: Improving Chain-of-Thought Reasoning in LLMs,” Advances in Neural Information Processing Systems , vol. 37, pp. 333–356, Dec. 2024

work page 2024

[7] [7]

Large Language Models Are Zero-Shot Reasoners,

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large Language Models Are Zero-Shot Reasoners,” Advances in Neural In- formation Processing Systems , vol. 35, pp. 22 199–22 213, Dec. 2022

work page 2022

[8] [8]

Automatic chain of thought prompting in large language models,

Z. Zhang, A. Zhang, M. Li, and A. Smola, “Automatic chain of thought prompting in large language models,”The Eleventh International Conference on Learning Representations , Kigali, Rwanda, May 2023

work page 2023