Context-Aware Semantic Communication for the Wireless Networks

Abbas Jamalipour; Dusit Niyato; Guangyuan Liu; Hongyang Du; Jiacheng Wang; Jiawen Kang; Yinqiu Liu; Zehui Xiong

REVIEW 6 major objections 5 minor 1 cited by

Context-Aware Semantic Communication for the Wireless Networks

T0 review · 6 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 CaSemCom uses an LLM context gate to choose which semantic features to transmit, improving image fidelity and cutting retransmissions.

desk verdict The architecture is reasonably clear and the context taxonomy is useful, but the case study doesn't actually measure bandwidth and the LLM gate is unspecified, so the central claim is still unverified. read the letter →

arxiv 2505.23249 v1 pith:ETRWAM6N submitted 2025-05-29 cs.NI eess.SP

Guangyuan Liu , Yinqiu Liu , Jiacheng Wang , Hongyang Du , Dusit Niyato , Jiawen Kang , Zehui Xiong , Abbas Jamalipour This is my paper

classification cs.NIeess.SP

keywords semanticcommunicationcontextawarenesslargelanguagemodelsmixtureofexpertswirelessnetworksdeepreinforcementlearningfidelitymultimodal

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

The paper argues that semantic communication should be context-aware: instead of always transmitting all features with one fixed encoder, a wireless system should decide, moment by moment, which parts of the data carry meaning for the current task and which expert encoder should process them. It proposes CaSemCom, in which a Large Language Model acts as a gate that reads task context (latency, fidelity needs) and communication context (SNR, bandwidth) and outputs content-selection and expert-selection decisions. A Mixture of Experts pool containing edge, pose, segmentation, depth, and text encoders then encodes only the chosen modalities, and the receiver reconstructs from the transmitted semantics plus lightweight control signals. In a multimodal, ten-user case study under Rayleigh fading, the paper reports that CaSemCom with the LLM gate converges faster, reaches higher reconstructed image quality (Image Reward plus SSIM), and incurs far fewer retransmissions than pure DQN, DeepSC, greedy, and random baselines, with a DRL fallback preserving most of the gain during LLM outages. A sympathetic reader would care because static semantic communication models waste bandwidth on irrelevant details, and this framework turns semantic selection into an adaptive, context-driven optimization problem.

What carries the argument

The central mechanism is the LLM-based agentic context gate coupled with a Mixture of Experts encoder pool. The gate ingests task context (application priority, latency tolerance, semantic fidelity threshold) and communication context (SNR, bandwidth, channel state information), then outputs two decisions: which input modalities to keep and which expert encoders to activate. The expert pool includes Canny edge detection, human pose estimation, image segmentation, depth mapping, and textual semantics, and only the selected experts transmit their compressed semantic representations, reducing bandwidth while preserving task-relevant meaning. The paper also introduces a DRL fallback: decisions made by the LLM are stored as state-action pairs, and a DQN agent trained on that experience pool reproduces gating choices when LLM inference is unavailable, simulated at a 20% outage probability in the case study.

What would settle it

Replace the LLM gate's content-selection output with a random choice of the same number of modalities under the same bandwidth, and compare the combined Image Reward and SSIM; if the random gate matches the LLM gate's quality, the claimed critical intelligence layer is not doing the work. A complementary check is to have human annotators mark which modalities are task-relevant on a held-out set and measure the LLM gate's top-1 agreement with those labels.

Watch

Extended reading notes

Core claim

CaSemCom's central claim is that an LLM-based gating mechanism can serve as the critical intelligence layer of a semantic communication system, performing two decisions that static models cannot: input content selection, which identifies the semantically salient portions of raw data (for example faces and speech in video conferencing, or bounding boxes in driving), and expert selection, which activates the specialized encoders that should process the chosen content. The gate takes task context and communication context, and its decisions, saved as state-action pairs, can train a DRL fallback that takes over when LLM inference is unavailable. The case study shows that this design outperforms single-agent DRL and non-adaptive baselines in convergence speed, semantic accuracy, and retransmission overhead, with CaSemCom with LLM reaching rewards above 0.8 within 50 steps and the DRL fallback around 0.75, while pure DQN converges near 0.4 and greedy and random schemes fluctuate around 0.3.

Load-bearing premise

The load-bearing premise is that the LLM-based context gate can reliably decide, without task-specific training, which semantic modalities (edge, pose, segmentation, depth, text) matter for the current task and channel state; the paper presents no direct evaluation of that selection quality.

Editorial extensions

If this is right

If the results hold, a single semantic communication stack can serve heterogeneous tasks by swapping active experts instead of retraining the whole encoder.
Bandwidth usage drops because only task-critical modalities are transmitted; in the case study this also translates into faster convergence of the learned gating policy.
The DRL fallback, trained from logged LLM decisions, keeps the system operational when the LLM is unavailable, with only a modest drop in reconstructed quality.
Under strict semantic fidelity thresholds, retransmission overhead becomes the differentiator, so context-aware gating matters most for reliable, low-latency applications.
The framework turns the LLM's gating decisions into a reusable experience pool, meaning fallback policies can be continuously distilled from LLM reasoning.

Reading between the lines

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

Editorial extension: the paper does not specify the LLM prompt, model version, or a direct measure of selection quality, so the strongest unsupported step is whether the gate's zero-shot modality choices are actually correct; a wrong gate would erase the reported gains.
Editorial extension: the case study only covers vision and text modalities under Rayleigh fading, but the same gate-plus-experts design should extend to audio and to longer textual inputs once chunking is added, as the paper itself notes.
Editorial inference: because the DRL fallback is trained on LLM decisions, it inherits whatever biases the LLM gate has; a defective gate would produce a confidently wrong fallback policy, not just a less accurate one.
Editorial note on the manuscript: Section IV-A3 lists hit rate and caching overhead as the performance indicators, while Section IV-B reports retransmission overhead; the stated metrics and the reported results are not reconciled.

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

The paper proposes CaSemCom, an LLM-based context gate on top of a MoE semantic encoder pool, and evaluates it in a multimodal multi-user simulation. The architecture itself is presented clearly: the context aggregator, the two-step gating decision (content selection then expert selection), and the DRL fallback are all easy to follow. The taxonomy in Section II — internal vs. external context across linguistic, wireless, and semantic layers — is a genuinely useful framing and is well grounded in examples like CarBeam and the LTC framework. That part is solid.

The soft spots are concentrated in the case study, and they are serious. The abstract claims improved fidelity while reducing bandwidth usage, but no bandwidth measurement appears anywhere. Figure 4 reports a combined Image Reward/SSIM score and Figure 5 reports retransmission counts. Retransmissions are a proxy for overhead, not a bit-level measure. The text in Section IV-A3 literally defines "Hit Rate" and "Caching Overhead" — leftover from a caching paper template — and never uses them. That looks like a copy-paste mistake, which does not inspire confidence in the numerical results.

The LLM gate, the core innovation, is also under-specified. Section III-B describes it as "the critical intelligence layer" but gives no model name, prompt template, temperature, or decision protocol. We cannot tell whether the gate is doing zero-shot reasoning, few-shot prompting, or effectively hand-coded rules. If the LLM's selections are arbitrary, the reported convergence speed and final reward are not reproducible. The DRL fallback, meanwhile, is trained on state-action pairs generated by LLM inference, so it is a distillation of the LLM's choices — comparing it to "pure DQN" is not a fair comparison of learned adaptation.

That said, these flaws are addressable. The idea of using an LLM to select semantic modalities based on task and channel context is worth exploring, and the authors have the components in their prior work. The paper would benefit from a real implementation specification for the gate, a direct measurement of transmitted bits or compression ratio, and a comparison against an independent learned policy. With those additions, the case study could support the claims.

For now, this is a concept paper with a plausible architecture and an under-built validation. I would send it to peer review only with the expectation of major revision — a referee could usefully force the authors to pin down the LLM gate and report actual bandwidth numbers. A desk reject would be too harsh, because the framing and taxonomy are worth preserving.

Recommendation: engage with it, but treat the numerical results as preliminary until the missing measurements are provided.

Referee Report

6 major / 5 minor

Summary. The manuscript proposes CaSemCom, a context-aware semantic communication framework in which an LLM-based 'agentic context gate' selects the semantically important segments of input content and the specialized semantic encoders (Mixture of Experts) to transmit, with a DRL fallback for cases when LLM inference is unavailable. After a taxonomy of context in linguistic, wireless, and semantic communication, the paper presents a multimodal multi-user simulation under Rayleigh fading comparing CaSemCom with LLM, CaSemCom with DRL fallback, pure DQN, DeepSC, greedy, and random baselines. The reported results show higher combined ImageReward/SSIM reward and lower retransmission counts for the proposed method.

Significance. The idea of an LLM-based context gate for adaptive semantic communication is timely, and the context taxonomy in Section II is a useful organizing framework. The paper correctly identifies a real limitation of static SemCom schemes, and the DRL fallback is a sensible robustness mechanism. However, the contributions are currently supported only by an under-specified simulation: the claimed bandwidth reduction is not measured, the LLM gate is not described at an implementable level of detail, and no variance information is given. If the framework is fully specified and the experiments are extended with bit-level accounting and multiple independent runs, the idea could become a meaningful contribution to adaptive semantic communication. At present, the empirical evidence is insufficient to verify the central claims.

major comments (6)

[§IV-B and abstract] The abstract claims that CaSemCom 'reduces bandwidth usage,' but the case study reports no bandwidth, compression-ratio, or transmitted-bit measurement: Fig. 4 plots a combined ImageReward/SSIM reward, and Fig. 5 plots retransmission counts, which are not bandwidth units. A retransmission counter does not quantify bandwidth consumption, so the bandwidth-reduction claim in the abstract is unsupported by the reported evidence.
[§III-B] The LLM-based context gate is the core novelty, but it is described only in prose. No LLM version, prompt template, temperature, output-parsing rule, or evaluation of gate-selection accuracy is provided. Without this information, the reader cannot reproduce the gate or exclude the possibility that it acts as an oracle, and the performance advantage attributed to the gate in Fig. 4 is not verifiable.
[§IV-B1 and §III-B3] The normalization of the combined ImageReward/SSIM reward to [0,1] is never explained. ImageReward can take values outside [0,1], and the 'predetermined weights' in §III-B3 are not reported; consequently, the reward values in Fig. 4 and the fidelity thresholds 0.3, 0.6, and 0.9 in Fig. 5 lack a clear quantitative meaning.
[§III-B3 and §IV-B1] The DRL fallback is trained on LLM-generated state-action pairs, so 'CaSemCom with DRL fallback' is a distillation of LLM decisions rather than an independent single-agent DRL method. Comparing it against a Pure DQN trained without demonstrations is not a matched comparison and does not support the claim of outperforming single-agent DRL; a DQN baseline with identical training data and reward would be needed.
[§IV-A3] The 'Performance Indicators' subsection defines Hit Rate and Caching Overhead for caching strategies, yet no caching experiment is reported and these metrics are never used in §IV-B. This is a leftover from a different paper template and should be removed or replaced with the metrics that actually support the paper's claims.
[§IV-B, Figs. 4 and 5] No error bars, confidence intervals, or multiple-seed results are reported. Since the text repeatedly asserts that CaSemCom 'significantly outperforms' baselines, the absence of variance information means that the claimed significance is not established; the authors should report the number of independent runs and a statistical comparison.

minor comments (5)

[§IV-B1] The number of training steps is inconsistently given as 250 in the Fig. 4 caption and 240 in the text; please reconcile.
[§I] In the contribution list, 'CaSeCom leverages' should be 'CaSemCom leverages,' and 'basline approaches' should be 'baseline approaches.'
[§IV-A2] The 'diffusion-based model' used for reconstruction is not identified; please specify the model and version.
[§IV-B1] The 'Simplified DeepSC [15]' baseline is not described in terms of which simplifications are applied; please provide the baseline implementation details.
[Fig. 5] The retransmission-overhead bars have numeric labels that are not explained in the caption; adding a legend or a tabular breakdown would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the case study is self-contained and the self-citations are not load-bearing.

full rationale

The paper's central claim is supported by an implemented, externally compared simulation, not by a derivation that reduces to its own inputs. No equation in the paper defines a predicted quantity in terms of the same quantity that is later reported as the result. The LLM-based gating and MoE expert selection are design choices; the case study measures reward and retransmission overhead against pure DQN, DeepSC, greedy, and random baselines, so the comparison has independent content. The DRL fallback is explicitly trained on past LLM-inferred decisions, but the paper presents this as a robustness mechanism for LLM outage rather than as a claim that the fallback's performance is predicted from the LLM; the measured curves are empirical. Self-citations such as [3] and [11] are used as inspiration and as a source for the five vision modalities, but they are not invoked as a uniqueness theorem or as the sole justification for the central result. The absence of a bandwidth-usage measurement and the unspecified LLM prompt/version are reproducibility and evidence gaps, not circularity. Therefore, no circular step can be exhibited, and the correct finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The framework rests on domain assumptions about the simulation, not on a mathematical derivation. The free parameters are hand-set simulation and training choices. The LLM gate and the multimodal experts are taken from the authors' prior work, and no independent falsifiable handle is provided for the gate. These assumptions are not necessarily wrong, but they are not validated outside the reported curves.

free parameters (4)

Reward weights for ImageReward and SSIM = Not specified
Section IV-A3 says the reward is a combined Image Reward and SSIM with predetermined weights, but the weights are never given, so the exact reward cannot be reproduced.
Semantic fidelity retransmission thresholds = 0.3, 0.6, 0.9
Section IV-B2 uses these thresholds to trigger retransmissions; they are chosen by hand and directly determine the overhead comparison.
DQN fallback hyperparameters = LR 1e-3, discount 0.995, replay buffer 2000, batch 32, epsilon 1.0 to 0.01, 250 episodes
Section IV-A2 lists these training choices with no sensitivity analysis; different values could change convergence and overhead results.
Wireless simulation parameters = 1.4 MHz bandwidth, 10 users, SNR range -13 to 30 dB
Section IV-A2 fixes these values; the reported advantage may depend on this specific channel regime.

assumptions (4)

domain assumption Rayleigh fading with fixed 1.4 MHz bandwidth and SNR range -13 to 30 dB is representative of realistic wireless SemCom deployments.
Invoked in Section IV-A2 to justify the case study; no comparison with other channel models or mobility patterns is provided.
domain assumption The combined ImageReward and SSIM metric, normalized to [0,1], is a valid proxy for semantic fidelity.
Used in Section IV-A3 and Figure 4; the normalization method is not described, and ImageReward is a learned human-preference model rather than a calibrated fidelity measure.
domain assumption The LLM context gate can select high-impact semantic modalities zero-shot, without task-specific training or fine-tuning.
Assumed in Section III-B; the paper provides no evaluation of selection accuracy and no ablation without the LLM.
domain assumption Five semantic modalities, edge, pose, segmentation, depth, and text, are sufficient to reconstruct high-quality images for the chosen prompts.
Introduced in Section IV-A1; no evidence is given that other modalities or combinations were needed or would change the results.

invented entities (1)

LLM-based agentic context gate
purpose: Selects task-critical input content and activates the appropriate semantic encoder experts in the CaSemCom framework.
The gate is the core architectural contribution but has no falsifiable prediction outside the paper's simulation and is not validated independently of the reported reward curves.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Context-Aware Semantic Communication for the Wireless Networks." pith.science (2026). https://pith.science/paper/ETRWAM6N

@misc{pith2026250523249,
  author       = {Pith},
  title        = {Pith review of: Context-Aware Semantic Communication for the Wireless Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETRWAM6N}},
  note         = {Machine review of arXiv:2505.23249}
}

read the original abstract

In next-generation wireless networks, supporting real-time applications such as augmented reality, autonomous driving, and immersive Metaverse services demands stringent constraints on bandwidth, latency, and reliability. Existing semantic communication (SemCom) approaches typically rely on static models, overlooking dynamic conditions and contextual cues vital for efficient transmission. To address these challenges, we propose CaSemCom, a context-aware SemCom framework that leverages a Large Language Model (LLM)-based gating mechanism and a Mixture of Experts (MoE) architecture to adaptively select and encode only high-impact semantic features across multiple data modalities. Our multimodal, multi-user case study demonstrates that CaSemCom significantly improves reconstructed image fidelity while reducing bandwidth usage, outperforming single-agent deep reinforcement learning (DRL) methods and traditional baselines in convergence speed, semantic accuracy, and retransmission overhead.

Figures

Figures reproduced from arXiv: 2505.23249 by the authors.

**Figure 1.** A unified illustration of linguistic communication, SemCom, and wireless transmission, each governed by internal [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Architecture of the proposed LLM-based CaSemCom framework. At the transmitter, the system determines which parts [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** An overview of the LLM-based context gate and its [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Evolution of the average reconstructed image quality, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Retransmission overhead (log scale) for six methods [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Wireless Copilot: An AI-Powered Partner for Navigating Next-Generation Wireless Complexity
cs.NI 2025-12 conditional novelty 5.0 of 10

Introduces a human-in-the-loop 'Wireless Copilot' framework for 6G network management, with a LAWNets simulation showing higher intent satisfaction than LLM/RL baselines.

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Semantic communications for future internet: Fundamentals, applications, and challenges,

W. Yang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for future internet: Fundamentals, applications, and challenges,” IEEE Communications Surveys & Tutorials , vol. 25, no. 1, pp. 213–250, 2022

work page 2022
[2]

Context-aware wireless sensor networks for assisted living and residential monitoring,

A. D. Wood, J. A. Stankovic, G. Virone, L. Selavo, Z. He, Q. Cao, T. Doan, Y . Wu, L. Fang, and R. Stoleru, “Context-aware wireless sensor networks for assisted living and residential monitoring,” IEEE network , vol. 22, no. 4, pp. 26–33, 2008

work page 2008
[3]

Mixture of experts for network optimization: A large language model- enabled approach,

H. Du, G. Liu, Y . Lin, D. Niyato, J. Kang, Z. Xiong, and D. I. Kim, “Mixture of experts for network optimization: A large language model- enabled approach,” arXiv preprint arXiv:2402.09756 , 2024

arXiv 2024
[4]

Adapting llm agents with universal feedback in communication,

K. Wang, Y . Lu, M. Santacroce, Y . Gong, C. Zhang et al. , “Adapting llm agents with universal feedback in communication,” in ICML 2024 Workshop on F oundation Models in the Wild , 2024

work page 2024
[5]

Context-aware network selection in heterogeneous wireless networks,

A. Monteiro, E. Souto, R. Pazzi, and M. Nogueira, “Context-aware network selection in heterogeneous wireless networks,” Computer Com- munications, vol. 135, pp. 1–15, 2019

work page 2019
[6]

Context-aware beam tracking for 5g mmwave v2i communications,

H. Ding and K. G. Shin, “Context-aware beam tracking for 5g mmwave v2i communications,” IEEE Transactions on Mobile Computing , vol. 22, no. 6, pp. 3257–3269, 2023

work page 2023
[7]

Extended context-based semantic communication system for text transmission,

Y . Liu, S. Jiang, Y . Zhang, K. Cao, L. Zhou, B.-C. Seet, H. Zhao, and J. Wei, “Extended context-based semantic communication system for text transmission,” Digital Communications and Networks , vol. 10, no. 3, pp. 568–576, 2024

work page 2024
[8]

Large ai model empowered multimodal semantic communications,

F. Jiang, L. Dong, Y . Peng, K. Wang, K. Yang, C. Pan, and X. You, “Large ai model empowered multimodal semantic communications,” IEEE Communications Magazine , vol. 63, no. 1, pp. 76–82, 2025

work page 2025

Show all 15 references

[9]

Toward mixture-of-experts enabled trustworthy semantic communication for 6g networks,

J. He, X. Luo, J. Kang, H. Du, Z. Xiong, C. Chen, D. Niyato, and X. Shen, “Toward mixture-of-experts enabled trustworthy semantic communication for 6g networks,” IEEE Network , 2024

2024
[10]

From semantic communication to semantic-aware networking: Model, architecture, and open problems,

G. Shi, Y . Xiao, Y . Li, and X. Xie, “From semantic communication to semantic-aware networking: Model, architecture, and open problems,” IEEE Communications Magazine , vol. 59, no. 8, pp. 44–50, 2021

2021
[11]

Semantic communications for artificial intelligence generated content (aigc) toward effective content creation,

G. Liu, H. Du, D. Niyato, J. Kang, Z. Xiong, D. I. Kim, and X. Shen, “Semantic communications for artificial intelligence generated content (aigc) toward effective content creation,” IEEE Network , 2024

2024
[12]

ImageReward: Learning and Evaluating Human Preferences for Text- to-Image Generation,

J. Xu, X. Liu, Y . Wu, Y . Tong, Q. Li, M. Ding, J. Tang, and Y . Dong, “ImageReward: Learning and Evaluating Human Preferences for Text- to-Image Generation,” Proc. Adv. Neural Inf. Process. Syst. , vol. 36, 2024

2024
[13]

On the Mathematical Properties of the Structural Similarity Index,

D. Brunet, E. R. Vrscay, and Z. Wang, “On the Mathematical Properties of the Structural Similarity Index,” IEEE Trans. Image Process., vol. 21, no. 4, pp. 1488–1499, 2011

2011
[14]

Dqn-based predictive spectrum handoff via hybrid priority queuing model,

H. Luo, K. Cao, Y . Wu, X. Xu, and Y . Zhou, “Dqn-based predictive spectrum handoff via hybrid priority queuing model,” IEEE Communi- cations Letters , vol. 26, no. 3, pp. 701–705, 2022

2022
[15]

Deep learning enabled semantic communication systems,

H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE transactions on signal pro- cessing, vol. 69, pp. 2663–2675, 2021

2021

Pith tools

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

[1] [1]

Semantic communications for future internet: Fundamentals, applications, and challenges,

W. Yang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for future internet: Fundamentals, applications, and challenges,” IEEE Communications Surveys & Tutorials , vol. 25, no. 1, pp. 213–250, 2022

work page 2022

[2] [2]

Context-aware wireless sensor networks for assisted living and residential monitoring,

A. D. Wood, J. A. Stankovic, G. Virone, L. Selavo, Z. He, Q. Cao, T. Doan, Y . Wu, L. Fang, and R. Stoleru, “Context-aware wireless sensor networks for assisted living and residential monitoring,” IEEE network , vol. 22, no. 4, pp. 26–33, 2008

work page 2008

[3] [3]

Mixture of experts for network optimization: A large language model- enabled approach,

H. Du, G. Liu, Y . Lin, D. Niyato, J. Kang, Z. Xiong, and D. I. Kim, “Mixture of experts for network optimization: A large language model- enabled approach,” arXiv preprint arXiv:2402.09756 , 2024

arXiv 2024

[4] [4]

Adapting llm agents with universal feedback in communication,

K. Wang, Y . Lu, M. Santacroce, Y . Gong, C. Zhang et al. , “Adapting llm agents with universal feedback in communication,” in ICML 2024 Workshop on F oundation Models in the Wild , 2024

work page 2024

[5] [5]

Context-aware network selection in heterogeneous wireless networks,

A. Monteiro, E. Souto, R. Pazzi, and M. Nogueira, “Context-aware network selection in heterogeneous wireless networks,” Computer Com- munications, vol. 135, pp. 1–15, 2019

work page 2019

[6] [6]

Context-aware beam tracking for 5g mmwave v2i communications,

H. Ding and K. G. Shin, “Context-aware beam tracking for 5g mmwave v2i communications,” IEEE Transactions on Mobile Computing , vol. 22, no. 6, pp. 3257–3269, 2023

work page 2023

[7] [7]

Extended context-based semantic communication system for text transmission,

Y . Liu, S. Jiang, Y . Zhang, K. Cao, L. Zhou, B.-C. Seet, H. Zhao, and J. Wei, “Extended context-based semantic communication system for text transmission,” Digital Communications and Networks , vol. 10, no. 3, pp. 568–576, 2024

work page 2024

[8] [8]

Large ai model empowered multimodal semantic communications,

F. Jiang, L. Dong, Y . Peng, K. Wang, K. Yang, C. Pan, and X. You, “Large ai model empowered multimodal semantic communications,” IEEE Communications Magazine , vol. 63, no. 1, pp. 76–82, 2025

work page 2025