Pith. sign in

REVIEW 3 major objections 49 references

Tiny language models can close the semantic gap in 6G by fitting meaning-first encoders onto constrained IoT and edge hardware under a latency-accuracy-size trilemma.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 03:49 UTC pith:SZDEUV72

load-bearing objection Useful 6G semantic/t-LM roadmap with a real taxonomy and open-problem list; the headline compression numbers and SES/Pareto synthesis overstate comparability across tasks. the 3 major comments →

arxiv 2607.03246 v1 pith:SZDEUV72 submitted 2026-07-03 eess.SP

Bridging the Semantic Gap in 6G: Tiny Language Models Under the Latency-Accuracy-Size Trilemma

Arnav Mathur , Garima Mathur , Rahul Jashvantbhai Pandya This is my paper
classification eess.SP
keywords Semantic Communication6G NetworksTiny Language ModelsModel CompressionEdge AIResource AllocationKnowledge GraphsJoint Source-Channel Coding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Sixth-generation networks are framed as AI-native systems that should carry meaning rather than raw bits, yet the deep models used for semantic encoding are far too large, slow, and power-hungry for the sensors and edge chips that will dominate deployments. This survey argues that tiny language models—compact, quantized, task-specialized models—are the practical bridge. It unifies semantic information theory with a two-axis taxonomy of five system architectures and six compression methods, then shows that pruning, quantization, distillation, LoRA, and especially split computing can shrink encoders by orders of magnitude while largely keeping task quality. Reported high-water marks include up to 99.98% size reduction, device-side models with only 640 parameters, and about 65% lower transmission energy when knowledge graphs guide what is sent. The paper then names open gaps in theory, security, knowledge-base sync, and hardware co-design, and sketches a 3GPP path toward IMT-2030.

Core claim

Across the surveyed systems, model compression can cut semantic-encoder size by as much as 99.98% while preserving task accuracy, split computing can leave only 640 parameters on the device, and knowledge-graph integration can reduce transmission energy by about 65%—evidence that tiny language models make semantic communication deployable on IoT- and edge-class 6G hardware under the latency-accuracy-size trilemma.

What carries the argument

The latency-accuracy-size trilemma, organized by a 5×6 taxonomy of architectures (end-to-end JSCC, split learning, federated learning, knowledge-graph-assisted, multi-task/cross-modal) against compression methods (quantization, pruning, knowledge distillation, LoRA, split computing, NAS), plus a Semantic Efficiency Score that normalizes task quality by log model size for cross-study comparison.

Load-bearing premise

That the mixed quality numbers from different studies—BLEU, PSNR, accuracy, and similar scores—can be fairly normalized into one ranking without selection or reporting bias in the chosen corpus.

What would settle it

Re-run the Pareto-frontier systems (extreme pruning, hierarchical progressive compression, split meta-learning, and federated distillation) on one shared hardware tier and one shared task suite; if extreme compression then collapses task accuracy far below the claimed retention, or if SES rankings reverse under a single common metric, the central deployability claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Semantic encoders can be matched to 6G slice classes: sub-kilobyte models for mMTC sensors, split encoders for URLLC, larger edge models for eMBB.
  • Neural architecture search aimed at semantic distortion remains an empty cell and is the largest unexplored design path.
  • LoRA and federated distillation make knowledge-base synchronization cheap enough for NB-IoT-class uplinks.
  • Standardization can sequence semantic KPIs, compressed KB exchange, and post-quantum-aware MACs from study items toward IMT-2030.
  • Hardware co-design and channel-adaptive compression are required to hit sub-millisecond semantic scoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If over-parameterization is general, future semantic codecs may be designed tiny from the start rather than compressed from large transformers.
  • SES will only become a useful community ranking tool if papers routinely report paired quality and model-size numbers under common channel conditions.
  • Post-quantum signature size may force hybrid authentication that applies full PQC only to high-value semantic sessions and lighter checks to routine telemetry.
  • Multi-cell networks with overlapping knowledge bases may need MAC policies that treat semantic correlation as a distinct interference type, not only power and beamforming.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. This survey argues that tiny language models (t-LMs) are the practical bridge for deploying semantic communication on resource-constrained 6G endpoints under a latency–accuracy–size trilemma. It synthesizes semantic information theory (entropy, capacity, rate–distortion/IB), proposes a 5×6 architecture×compression taxonomy, reviews quantisation, pruning, KD, LoRA, split computing, and NAS through a semantic-quality lens, and surveys multi-user resource allocation and KB management. Headline quantitative claims—up to 99.98% encoder compression with retained task accuracy, device-side encoders with 640 parameters, and 65% energy reduction via knowledge graphs—are drawn from primary studies and re-aggregated via a Semantic Efficiency Score (SES, Eq. 25), Pareto analysis (Fig. 3), and Cohen’s d forest plot. Seven (later nine) open challenges and a 3GPP/IMT-2030 roadmap complete the contribution.

Significance. If the synthesis is accepted as a fair map of the literature, the paper is a useful organizing contribution for eess.SP and 6G systems: it connects semantic information theory to edge-deployable model compression, makes the empty NAS×semantic-communication cell explicit, and packages concrete deployment numbers (Lite-DeepSC 40×, HAPE 63×, SPM 1 kB / 99.98% FLOP, Semantic-MSL 640 params, KG 65% energy, FD 25.6× comm) with a standardisation timeline. Strengths include the two-axis taxonomy (Fig. 1, Table IV), the structured evidence table (Table VI), and an explicit open-problem list spanning capacity, sub-ms scoring, PQC for IoT, and KB sync. The new SES/Pareto layer is a genuine attempt at cross-study comparison rather than pure catalogue, but its validity is the main load-bearing risk for the quantitative narrative.

major comments (3)
  1. Sections VIII–IX and Eq. (25)/Fig. 3: The abstract and strongest claim present 99.98% compression, 640 device-side parameters, and 65% energy cut as joint evidence that t-LM compression closes the trilemma. Those figures come from different primary tasks and hardware (SPM binary classification on a 1 kB MCU; Semantic-MSL few-shot image classification with edge offload; KG triple selection energy). F_norm is quality relative to each study’s own full-model baseline, and SES is defined only when both quality and size are reported (14 of the surveyed studies). The Pareto “shallow negative slope” and ranking can therefore be driven by task narrowness and reporting selection rather than a common semantic-quality notion. Please either (i) restrict the meta-analysis to within-modality, within-task cohorts with explicit inclusion criteria and sensitivity checks on baseline choice, or (ii) reframe
  2. §IX.E / Fig. 4: Cohen’s d is computed with pooled σ “estimated from reported test-set variance,” but most primary studies do not publish instance-level variance suitable for a common effect-size synthesis. Without a transparent protocol for σ estimation (or a switch to simple relative degradation with reported ranges), the forest plot overstates statistical precision. Either document the estimation procedure study-by-study or replace d with relative quality change and confidence bands only where the source papers support them.
  3. Abstract vs. §X: The abstract states “Seven open challenges,” while §X formalises nine (C1–C9), and Table IX maps seven. Align the count and labels throughout, and ensure the abstract’s challenge list matches the body so the contribution claim is consistent.

Circularity Check

0 steps flagged

No significant circularity: a literature survey reporting external primary results, with SES/Pareto as post-hoc comparative constructs that do not force the headline compression numbers by construction.

full rationale

This is a structured review of external work on t-LM semantic communication for 6G. The load-bearing quantitative claims (99.98% FLOP/size reduction with retained accuracy [1/SPM], 640 device-side parameters [2/Semantic-MSL], 65% transmission-energy cut [3/KG]) are attributed to cited primary studies by other authors and are not derived from parameters fitted in this paper. The authors’ own constructs—SES in (25), F_norm normalization, the 5×6 taxonomy, and the Pareto plot in Fig. 3—are comparative aggregation tools applied after the fact to studies that already report quality and size; they do not redefine those primary numbers or turn a fit into a “prediction.” There is no self-definitional loop (X defined via Y then used to derive Y), no fitted-input-called-prediction, no load-bearing uniqueness theorem or ansatz imported from overlapping authors, and no renaming of a known law as a new derivation. Heterogeneous metrics and post-hoc ranking raise comparability/selection concerns for the synthesis narrative, but those are validity issues, not circularity of a derivation chain. Score 0 with empty steps is therefore the correct outcome.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

As a survey the load-bearing content is imported from Shannon/Weaver, Bao et al. semantic entropy/capacity, IB formulations, and the primary experimental papers. The authors add definitional constructs (SES, taxonomy cells, S-SNR) whose numerical rankings depend on how they normalize heterogeneous metrics.

free parameters (3)
  • SES denominator log10(size/1 kB)
    Arbitrary reference size and log base chosen by the authors to produce a scalar ranking; different choices reorder systems.
  • F_norm baseline choice
    Normalisation of BLEU/accuracy/PSNR to each study’s own full-model baseline; not a universal scale.
  • Cohen’s d pooled σ estimates
    Forest-plot error bars are estimated from reported test-set variance rather than raw data.
axioms (4)
  • domain assumption Semantic channel capacity Cs can exceed Shannon capacity when receiver interpretive ability outweighs coding ambiguity (Bao et al. formulation).
    Used as theoretical foundation in §III; not re-derived.
  • domain assumption Task metrics (BLEU, PSNR, classification accuracy, MSS) are adequate proxies for semantic fidelity under compression.
    Underpins all RQ2/RQ3 claims and SES; paper itself notes lack of a universal metric (Challenge 5).
  • domain assumption Cited primary-study numbers (compression ratios, energy, latency) are accurate and comparable across hardware and channels.
    Required for Tables VI–VIII and Pareto frontier.
  • standard math Standard information-bottleneck and rate-distortion extensions apply to neural semantic encoders.
    Eqs. 7–9 imported from Tishby/Liu/Sana without new proof.
invented entities (3)
  • Semantic Efficiency Score (SES) no independent evidence
    purpose: Cross-study scalar ranking of quality per model-size overhead.
    Defined in Eq. 25; no external validation or standardisation.
  • Two-axis 5×6 taxonomy of t-LM semantic systems no independent evidence
    purpose: Organise architecture × compression research cells and highlight NAS gap.
    Author-constructed maturity grid (Fig. 1, Table IV).
  • Semantic noise / S-SNR model no independent evidence
    purpose: Separate physical and semantic distortion (Eqs. 13–14).
    Adopted/adapted from Qin et al.; paper notes it is unimplemented in surveyed systems.

pith-pipeline@v1.1.0-grok45 · 34073 in / 2822 out tokens · 32669 ms · 2026-07-12T03:49:57.828191+00:00 · methodology

0 comments
read the original abstract

Sixth-generation (6G) wireless networks are expected to serve as AI-native infrastructure, transmitting meaning rather than mere bits -- a shift that makes semantic communication the central paradigm for next-generation connectivity. Deep learning-based semantic encoders show compelling gains in bandwidth efficiency; however, their dependence on large transformer models with hundreds of millions of parameters is at odds with the sub-millisecond latency, microjoule energy budgets, and kilobyte memory footprints of the constrained IoT and edge devices that will dominate 6G endpoints. Tiny language models (t-LMs) -- compact, quantised, task-specialised models deployable on microcontrollers, mobile system-on-chips, and edge accelerators -- are the enabling technology for closing this gap. This review provides a unified treatment of (i) the theoretical foundations of semantic information, covering semantic entropy, channel capacity, and rate-distortion theory; (ii) a two-axis taxonomy of t-LM-based semantic communication systems across five architecture classes and six compression paradigms; (iii) a survey of model compression techniques -- quantisation, pruning, knowledge distillation, low-rank adaptation, split computing, and neural architecture search -- through the lens of semantic quality preservation; and (iv) semantic-aware resource allocation frameworks for 6G multi-user networks. Evidence across the surveyed literature shows that compression can reduce semantic encoder size by up to 99.98% while preserving task accuracy, that split computing achieves device-side encoders with as few as 640 parameters, and that knowledge graph integration cuts transmission energy by 65%. Seven open challenges are identified, spanning theoretical gaps, system design, knowledge-base management, post-quantum security, and hardware co-design, with a 3GPP standardisation roadmap toward IMT-2030.

Figures

Figures reproduced from arXiv: 2607.03246 by Arnav Mathur, Garima Mathur, Rahul Jashvantbhai Pandya.

Figure 1
Figure 1. Figure 1: Two-axis taxonomy of t-LM-based semantic communication for 6G. Rows denote system architecture; columns denote compression technique. Fill intensity indicates research maturity of each research cell. Split Computing. The device-side computation remains with layers 1 . . . k; the edge server handles layers k+1 . . . N. The compression ratio is: ρ = size(t − LM) size(LLM) , (12) where “size” refers to parame… view at source ↗
Figure 1
Figure 1. Figure 1: G. Imitation Learning-Based Semantic Reasoning A complementary approach to compression for achieving t-LM-scale semantic encoders is to distil not the weights but the reasoning behaviour of a large semantic model. Imitation learning trains the student encoder to replicate the action sequence of a teacher oracle (a large LLM) on a set of semantic communication tasks, using behavioural cloning: LIL = −E(x,a∗… view at source ↗
Figure 2
Figure 2. Figure 2: Four principal architecture paradigms for [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pareto frontier of compression ratio vs. normalised [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Forest plot of compression-induced semantic quality [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Normalised semantic spectral efficiency (SSE) per [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: End-to-end latency breakdown for four system con [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 3GPP standardisation roadmap for t-LM-enabled 6G semantic communication, from Release 18 AI/ML study items through IMT-2030 native deployment. Upper lane: 3GPP milestones. Lower lane: corresponding t-LM research directions from this survey. XII. CONCLUSION This survey has provided a comprehensive structured review of tiny language models for 6G semantic communication, synthesising evidence across the surve… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 13 linked inside Pith

  1. [1]

    K. Chen, Z. Qin, G. Y . Li, and B.-H. Juang. Semantic pruning model for ultra-compact semantic encoders on microcontrollers in 6G industrial IoT networks.IEEE Internet of Things Journal, 11(9):15922–15935, 2024

    2024

  2. [2]

    Eldeeb, M

    E. Eldeeb, M. Shehab, H. Alves, and M.-S. Alouini. Semantic meta-split learning: A TinyML scheme for few-shot wireless image classification. IEEE Transactions on Machine Learning in Communications and Net- working, 3:594–610, April 2025

    2025

  3. [3]

    Y . Wang, Z. Ma, S. Guo, Q. Gao, and X. Chen. Knowledge graph probability graph for energy-efficient semantic communication in 6G edge networks.IEEE Transactions on Green Communications and Networking, 8(2):712–725, 2024

    2024

  4. [4]

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li. Semantic communications: Principles and challenges. arXiv preprint arXiv:2201.01389, June 2022

    Pith/arXiv arXiv 2022

  5. [5]

    E. C. Strinati and S. Barbarossa. 6G networks: Beyond Shannon to- wards semantic and goal-oriented communications.Computer Networks, 190:107930, May 2021

    2021

  6. [6]

    Nguyen, T.-D

    T.-H. Nguyen, T.-D. Tran, H.-L. Vu, and T.-N. Do. Post-quantum cryp- tography integration for tiny language model-based 6G semantic com- munication networks.IEEE Internet of Things Journal, 11(18):29475– 29488, 2024. 21

    2024

  7. [7]

    C. E. Shannon and W. Weaver.The Mathematical Theory of Communi- cation. University of Illinois Press, Urbana, IL, USA, 1949

    1949

  8. [8]

    C. E. Shannon. A mathematical theory of communication.Bell System Technical Journal, 27(3):379–423, July 1948

    1948

  9. [9]

    Carnap and Y

    R. Carnap and Y . Bar-Hillel. An outline of a theory of semantic information. Technical Report 247, Research Laboratory of Electronics, MIT, Cambridge, MA, USA, October 1952

    1952

  10. [10]

    J. Bao, P. Basu, M. Dean, C. Partridge, A. Swami, W. Leland, and J. A. Hendler. Towards a theory of semantic communication. InProc. IEEE Network Science Workshop (NetSciW), pages 110–117, West Point, NY , USA, June 2011

    2011

  11. [11]

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang. Deep learning enabled seman- tic communication systems.IEEE Transactions on Signal Processing, 69:2663–2675, April 2021

    2021

  12. [12]

    Bourtsoulatze, D

    E. Bourtsoulatze, D. B. Kurka, and D. Gündüz. Deep joint source- channel coding for wireless image transmission.IEEE Transactions on Cognitive Communications and Networking, 5(3):567–579, May 2019

    2019

  13. [13]

    D. B. Kurka and D. Gündüz. DeepJSCC-f: Deep joint source-channel coding of images with feedback.IEEE Journal on Selected Areas in Information Theory, 1(1):178–193, May 2020

    2020

  14. [14]

    Weng and Z

    Z. Weng and Z. Qin. Semantic communication systems for speech transmission.IEEE Journal on Selected Areas in Communications, 39(8):2434–2444, August 2021

    2021

  15. [15]

    H. Xie, Z. Qin, and G. Y . Li. Task-oriented multi-user semantic commu- nications for visual question answering.IEEE Wireless Communications Letters, 11(3):553–557, March 2022

    2022

  16. [16]

    Jiang, C.-K

    P. Jiang, C.-K. Wen, S. Jin, and G. Y . Li. Wireless semantic communi- cations for video conferencing. arXiv preprint arXiv:2204.07790, April 2022

    Pith/arXiv arXiv 2022

  17. [17]

    Luo, H.-H

    X. Luo, H.-H. Chen, and Q. Guo. Semantic communications: Overview, open issues, and future research directions.IEEE Wireless Communica- tions, 29(1):210–219, February 2022

    2022

  18. [18]

    Abdin, S

    M. Abdin, S. A. Jacobs, A. A. Awan, J. Aneja, A. Awadallah, H. Awadalla, N. Bach, A. Bahree, A. Bakhtiari, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, April 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Gemini: A family of highly capable multimodal models

    Gemini Team, Google. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, December 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Dubey, A

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, July 2024

    Pith/arXiv arXiv 2024

  21. [21]

    J. Liu, W. Zhang, and H. V . Poor. A rate-distortion framework for characterizing semantic information. arXiv preprint arXiv:2105.04278, May 2021

    Pith/arXiv arXiv 2021

  22. [22]

    Tishby, F

    N. Tishby, F. C. Pereira, and W. Bialek. The information bottleneck method. arXiv preprint arXiv:physics/0004057, 2000. Submitted to 37th Annual Allerton Conference

    Pith/arXiv arXiv 2000

  23. [23]

    Cheng, Z

    X. Cheng, Z. Zhang, L. Zhao, H. Zhang, and C. X. Wang. A com- prehensive review of AI-native 6G: Integrating semantic communica- tions, reconfigurable intelligent surfaces, and edge intelligence for next- generation connectivity.Frontiers in Communications and Networks, 6:1512843, 2025

    2025

  24. [24]

    S. R. Pokhrel, J. Choi, and W. Bennis. Urgency-based transmission scheduling for tiny language model-enabled semantic MAC layers in 6G networks.IEEE Transactions on Vehicular Technology, 72(10):13847– 13860, 2023

    2023

  25. [25]

    Agustsson, M

    E. Agustsson, M. Tschannen, F. Mentzer, R. Timofte, and L. V . Gool. Generative adversarial networks for extreme learned image compression. InProc. IEEE International Conference on Computer Vision (ICCV), pages 221–231, Seoul, South Korea, October 2019

    2019

  26. [26]

    Johnson, R

    J. Johnson, R. Krishna, M. Stark, L.-J. Li, D. A. Shamma, M. S. Bernstein, and L. Fei-Fei. Image retrieval using scene graphs. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3668–3678, Boston, MA, USA, June 2015

    2015

  27. [27]

    R. Hu, Y . Guo, H. Li, Q. Pei, and Y . Gong. Green federated learning for semantic communication via gradient sparsification in 6G IoT networks. IEEE Internet of Things Journal, 11(1):584–597, 2024

    2024

  28. [28]

    X. Liu, W. Jia, W. Liu, and W. Pedrycz. AFSSE: An interpretable clas- sifier with axiomatic fuzzy set and semantic entropy.IEEE Transactions on Fuzzy Systems, 28(11):2825–2840, November 2020

    2020

  29. [29]

    Sana and E

    M. Sana and E. C. Strinati. Learning semantics: An opportunity for effective 6G communications. arXiv preprint arXiv:2110.08049, October 2021

    Pith/arXiv arXiv 2021

  30. [30]

    Y . Lin, G. Zhu, Z. Zhang, D. Niyato, and X. Wang. Federated distillation for semantic communication: Communication-efficient knowledge shar- ing in multi-device 6G edge networks.IEEE Transactions on Wireless Communications, 22(12):9271–9285, 2023

    2023

  31. [31]

    Zhang, Q

    G. Zhang, Q. Hu, Z. Qin, Y . Cai, and G. Yu. A unified multi-task semantic communication system with domain adaptation. arXiv preprint arXiv:2206.00254, June 2022

    Pith/arXiv arXiv 2022

  32. [32]

    Zhang, J

    D. Zhang, J. Yang, D. Ye, and G. Hua. LQ-Nets: Learned quantization for highly accurate and compact deep neural networks. InProc. European Conference on Computer Vision (ECCV), pages 365–382, Munich, Germany, September 2018

    2018

  33. [33]

    Xie and Z

    H. Xie and Z. Qin. A lite distributed semantic communication system for internet of things.IEEE Journal on Selected Areas in Communications, 39(1):142–153, January 2021

    2021

  34. [34]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. InProc. International Conference on Learning Representations (ICLR), 2022

    2022

  35. [35]

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam. MobileNets: Efficient convo- lutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, April 2017

    Pith/arXiv arXiv 2017

  36. [36]

    Jiang, C.-K

    P. Jiang, C.-K. Wen, S. Jin, and G. Y . Li. Hierarchical attention- based progressive edge compression for semantic communications in resource-constrained 6G networks.IEEE Transactions on Wireless Communications, 22(11):7821–7836, 2023

    2023

  37. [37]

    Y . Liu, Z. Qin, X. Tao, and G. Y . Li. CCTaEncoder: A compact cross- task semantic encoder for multi-task wireless image communication. In Proc. IEEE Global Communications Conference (GLOBECOM), pages 1–6, Kuala Lumpur, Malaysia, 2023

    2023

  38. [38]

    J. Shi, S. Gao, Z. Qin, and G. Y . Li. FedPun: Federated pruning for semantic communication under non-IID data distributions. InProc. IEEE International Conference on Communications (ICC), pages 1–6, Rome, Italy, 2023

    2023

  39. [39]

    Y . Sun, Z. Qin, G. Y . Li, and P. Popovski. Synchronizing LLM-based semantic knowledge bases via secure federated fine-tuning in semantic communication.Frontiers in Artificial Intelligence, 8:1518782, 2025

    2025

  40. [40]

    Y . Sun, L. Yang, H. Zhao, B. He, and Q. Liu. FFA-LoRA: Freezing the first matrix A in LoRA for efficient fine-tuning of large language models. arXiv preprint arXiv:2403.18720, March 2024

    Pith/arXiv arXiv 2024

  41. [41]

    S. Qian, Z. Qin, G. Y . Li, and X. Tao. Contrastive disentanglement for compact semantic feature vectors in wireless communication systems. IEEE Transactions on Signal Processing, 72:1850–1864, 2024

    2024

  42. [42]

    H. Wu, X. Luo, T. Q. S. Quek, and H.-H. Chen. Energy-optimal cut-layer selection for split computing in tiny language model-based semantic communication systems.IEEE Wireless Communications Letters, 13(4):987–991, 2024

    2024

  43. [43]

    Y . Zhou, Z. Qin, and G. Y . Li. Adaptive reinforcement learning for knowledge-graph-assisted semantic question answering over wireless networks. InProc. IEEE Wireless Communications and Networking Conference (WCNC), pages 1–6, Glasgow, UK, 2023

    2023

  44. [44]

    Zhang, Z

    H. Zhang, Z. Kyaw, S.-F. Chang, and T.-S. Chua. Visual translation embedding network for visual relation detection. InProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5532–5540, Honolulu, HI, USA, July 2017

    2017

  45. [45]

    L. Wu, K. Huang, and H. Shen. A GAN-based tunable image com- pression system. InProc. IEEE International Conference on Computer Vision Workshops (ICCVW), pages 2334–2342, 2020

    2020

  46. [46]

    H. Tong, Z. Yang, S. Wang, Y . Hu, O. Semiari, W. Saad, and C. Yin. Federated learning for audio semantic communication.Frontiers in Communications and Networks, 2:43, September 2021

    2021

  47. [47]

    Zhang, Z

    X. Zhang, Z. Ma, Y . Wang, and S. Guo. Covert semantic communication via knowledge graph-based feature obfuscation for 6G networks. In Proc. IEEE International Conference on Communications (ICC), pages 1–6, Rome, Italy, 2023

    2023

  48. [48]

    Wang and S

    Y . Wang and S. Guo. Transceiver cooperative learning-aided seman- tic communications against mismatched background knowledge bases. arXiv preprint arXiv:2301.03133, January 2023

    Pith/arXiv arXiv 2023

  49. [49]

    L. Yan, Z. Qin, R. Zhang, Y . Li, and G. Y . Li. Resource allocation for semantic-aware networks. arXiv preprint arXiv:2201.06023, April 2022

    Pith/arXiv arXiv 2022