Pith. sign in

REVIEW 5 major objections 7 minor 55 references

TOAST: Task-Oriented Adaptive Semantic Transmission over Dynamic Wireless Environments

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TOAST fuses RL, diffusion, and LoRA into one adaptive semantic transmission framework.

desk verdict Plausible integration of known components, but the EDM denoiser's noise-level estimate uses transmitter-side clean data, so the central claim is unsupported as written. read the letter →

arxiv 2506.21900 v1 pith:VXIK4KZE submitted 2025-06-27 cs.LG eess.IV

Sheng Yun , Jianhua Pei , Ping Wang This is my paper · ORCID
classification cs.LGeess.IV
keywords task-orientedsemanticcommunicationjointsource-channelcodingSwinTransformerElucidatingdiffusionmodelreinforcementlearninglow-rankadaptationwirelessimagetransmissionmulti-taskbalancing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a single semantic communication system can adapt to changing wireless channels without full retraining, by combining three mechanisms: a reinforcement-learning agent that continuously reweights reconstruction fidelity against classification accuracy, module-specific LoRA adapters that specialize each component to a channel type, and an Elucidating Diffusion Model that denoises the received latent code. If the claim holds, a task-oriented transmitter can prioritize task-relevant information over bit-perfect delivery and still keep both image quality and classification usable in low-SNR regimes. The paper reports that on SVHN this integrated design beats the Swin-Transformer-only joint source-channel coding baseline by 8.4 dB PSNR and 9.8% accuracy at 5 dB SNR, and that LoRA adapters recover most performance on Rayleigh, Rician, phase-noise, and impulse-noise channels using 1% of the data and about 45 times fewer parameters than full fine-tuning.

What carries the argument

The load-bearing machinery is the layered pipeline in Algorithm 1. A Swin Transformer encoder maps the image to a power-normalized latent $z_{norm}$; the channel produces $z_{ch}$; the EDM denoiser, a continuous-time diffusion model with variance-preserving preconditioning, estimates the noise level as $\sigma_{\max} = \|z_{ch} - z_{norm}\|_2$ and iteratively refines $z_{ch}$ into a denoised latent; and a dual head reconstructs the image and classifies it. Around this pipeline sit two adaptors: a deep Q-network whose state is $[SNR, L_{recon}, Acc, epoch, \lambda_{prev}]$ and whose action picks $\lambda_{recon}$ with $\lambda_{recon}+\lambda_{cls}=1$, and LoRA updates $W' = W + \alpha_c B_c A_c$ with per-module ranks (encoder 16, decoder 16, EDM 8, classifier 4). The argument is that the RL weight scheduling, LoRA specialization, and EDM denoising each fix a separate failure mode of fixed-weight JSCC systems.

What would settle it

Run TOAST inference with only the received latent $z_{ch}$ available and test whether the receiver can produce a usable $\sigma_{\max}$ estimate without oracle access to $z_{norm}$; if no pilot, CSI-based, or learned estimator can supply it, Algorithm 1's EDM denoising loop cannot execute as written and the reported low-SNR gains would not be reproducible in deployment.

Watch

Extended reading notes

Core claim

The central claim is that TOAST, a framework uniting a Swin Transformer joint source-channel coding backbone, latent-space Elucidating Diffusion denoising, deep reinforcement-learning task balancing, and module-specific Low-Rank Adaptation, is the first system to jointly solve the multi-task, channel-adaptive, parameter-efficient semantic transmission problem. The paper argues that these components are complementary: the RL agent selects loss weights according to SNR and training progress, LoRA lets each module adapt cheaply to a specific impairment, and the EDM restores features corrupted by channel noise before both the reconstruction decoder and the classifier see them. Its experiments report consistent gains over JSCC-only and diffusion-augmented baselines across SVHN, CIFAR-10, Intel Image, and MNIST, with the largest improvements in the 0-10 dB SNR range.

Load-bearing premise

The receiver can estimate the noise level $\sigma_{\max}$ that the channel added to the latent code, even though the paper's formula for it uses the clean transmitted latent $z_{norm}$, which is not available at the receiver during inference.

Editorial extensions

If this is right

  • At 5 dB SNR on SVHN, the full TOAST pipeline is claimed to deliver 23.7 dB PSNR and 65.0% accuracy versus 15.3 dB and 55.2% for the Swin JSCC baseline, so the decisive gains sit exactly in the low-SNR regime 6G edge links face.
  • A new channel impairment can be handled by fine-tuning only the LoRA adapters (about 798.7K of 35.99M parameters) on 1% of data for up to five epochs, with reported gains on Rayleigh fading rising from 28.92% to 68.45% accuracy at 10 dB SNR.
  • The RL controller is expected to shift emphasis from reconstruction to classification as SNR improves, which would remove manual weight scheduling during deployment.
  • Because the task head and reward can be swapped, the same pipeline architecture extends to detection, segmentation, or multimodal tasks without changing the transmission core.

Reading between the lines

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

  • The reported gains of the EDM component are conditioned on knowing the clean transmitted latent to set $\sigma_{\max}$; a practical system would need a pilot-based or learned noise estimator, and the size of the gains under that estimator is an open question the paper does not test.
  • The paper's novelty is integration rather than mechanism: LoRA, EDM, and RL are established tools, so a sharper test of the framework would compare it against a strong fixed-weight JSCC with the same backbone and denoiser but no RL, which the current ablation does not fully isolate.
  • The adaptation experiments use 1% data sampled from the same distribution as training; a stiffer test would be adaptation to a genuinely new domain or channel type not present in any pretraining, which would show whether the LoRA library scales.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 7 minor

Summary. The paper proposes TOAST, a task-oriented semantic transmission framework that combines a Swin Transformer JSCC backbone, an Elucidating Diffusion Model (EDM) for latent-space denoising, a deep Q-network for adaptive weighting between reconstruction and classification losses, and module-specific LoRA adapters for channel-specific fine-tuning. The authors claim that TOAST achieves superior reconstruction quality and classification accuracy at low SNR compared to several baselines, and that it is the first framework to jointly address task balancing, channel adaptivity, parameter efficiency, and generative quality enhancement. Experiments are reported on SVHN, CIFAR-10, Intel Image, and MNIST under AWGN, Rayleigh, Rician, phase noise, and impulse noise channels.

Significance. If the reported results are reliable, the paper would make a useful contribution to task-oriented semantic communication by integrating several recently proposed components into one framework: RL-driven loss-weight scheduling, diffusion-based denoising, and parameter-efficient LoRA adaptation. The paper also provides a broad survey of 38 prior frameworks, which is helpful context. However, the central empirical claim is currently not dependable because the inference pipeline as described requires privileged access to the clean transmitted latent, and because there are internal inconsistencies in the reported EDM gains. The paper does not provide code, error bars, or comparisons to the most relevant published baselines, which further limits the significance of the experimental comparisons as presented.

major comments (5)
  1. [Algorithm 1, line 9] The EDM noise-level estimate is defined as sigma_max = ||z_ch - z_norm||_2, where z_norm is the clean, power-normalized latent at the transmitter. At the receiver only z_ch is available; z_norm is transmitted over the channel and is never reconstructed or otherwise estimated. Since Section III-C4 explicitly states that Algorithm 1 details the inference workflow, the EDM denoising loop in lines 10-12 cannot be executed as written. The paper provides no pilot scheme, CSI-based estimator, or fallback for estimating sigma_max at the receiver. This directly undermines the reported EDM gains in Table IV and Fig. 6, because the denoiser may be using information not available in the intended deployment. Please specify a receiver-side noise-level estimator, replace line 9 with an implementable rule, and re-run the experiments using that estimator.
  2. [Section VI.D and Section VI.E] The two quantitative statements about the EDM contribution are contradictory. Table IV shows that JSCC+EDM improves over JSCC-only by about 5-6 dB in PSNR on SVHN (e.g., 15.3 to 20.7 dB at 5 dB SNR), whereas Section VI.E reports that EDM contributes approximately 1.2 dB at 0 dB and nearly 1.5 dB at 5 dB SNR. Both cannot be true for the same experimental setup. This inconsistency affects the central ablation claim and must be resolved, either by correcting the reported numbers or by explaining what differs between the two comparisons.
  3. [Section IV.B] The paper describes a Deep Q-Network with a continuous action space and a Softplus output activation, but standard DQN is designed for discrete action spaces. No discretization of the weights is described, and no actor-critic alternative such as DDPG or TD3 is presented. Since adaptive task balancing is one of the three main contributions, the RL method must be specified precisely enough to be reproducible and technically sound. Please clarify how the continuous action constraint lambda_recon + lambda_cls = 1 is enforced and whether the method is actually DQN with a regression output or a different RL algorithm.
  4. [Section V.C and V.E] The LoRA adaptation strategy relies on an 'automatic channel-type detection module' that triggers adapter activation during inference, but no detection algorithm, training procedure, or detection accuracy is reported. Because the claim of adapting to unseen channel types depends on correctly selecting the channel-specific adapter, this missing component is load-bearing. Please specify how channel type is detected at the receiver, how detection errors affect performance, and how the adapter library is constructed.
  5. [Section VI.B and VI.D] The experimental comparison uses self-constructed baselines only: a CNN-based JSCC, a CNN with DDPM, and a 'Swin Transformer' baseline that is described as the authors' JSCC-only model. There is no comparison to the most relevant published systems, such as SwinJSCC [7], CDDM [17], or Diff-JSCC [8], even though these are cited in the introduction. In addition, no error bars, number of random seeds, or statistical significance tests are reported, and no code is provided. Given that the central claim is empirical superiority, the comparison needs to be placed in the context of existing published methods and accompanied by variance estimates.
minor comments (7)
  1. [Section II.A.1] The word 'task-orineted' appears to be a typo for 'task-oriented'.
  2. [Section III.C.3] The phrase 'presenters an in-depth exposition' should read 'presents an in-depth exposition'.
  3. [Section III.C.1] 'an Multilayer Perceptron classifier' should be 'a Multilayer Perceptron classifier'.
  4. [Section V.E] 'logit ouputs' should be 'logit outputs'.
  5. [Algorithm 1] The algorithm is introduced as an inference workflow, but lines 17-20 perform loss computation, RL policy updates, and LoRA adapter updates. Please clarify whether this is a training algorithm or an inference algorithm, and align the pseudocode with the text.
  6. [Table II] The adaptation time comparison (2.5 hours vs. 3 minutes) is reported without hardware details or training configuration; adding these would help reproducibility.
  7. [Section VI.D] The MNIST row shows 34.2 dB PSNR at 5 dB SNR for TOAST, which is higher than the PSNR of the clean input if the input is normalized to [0,1]; please verify these numbers and the compression setup.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: TOAST's gains are empirical comparisons against external baselines, and the self-citations are not load-bearing.

full rationale

The paper's central claims are empirical measurements of trained systems compared with implemented baselines, not quantities derived by construction from fitted parameters. Table III and Table IV report PSNR/accuracy for TOAST against JSCC-only and JSCC+EDM variants; these are run results, not predictions forced by the loss weights or LoRA ranks. The RL weight scheduler, LoRA adapters, and EDM denoiser are evaluated experimentally, with the EDM anchored to the external Karras et al. formulation [13] and LoRA to Hu et al. [50]. Self-citations [47], [48], and [52] appear in related-work motivation and efficiency arguments, but the core architecture does not reduce to those citations. The most serious issue is not circularity but an implementation gap: Algorithm 1 line 9 sets sigma_max = ||z_ch - z_norm||_2, while z_norm is the clean power-normalized latent known only at the transmitter; the receiver cannot compute this quantity, so the EDM denoising loop may rely on oracle noise-level information in the reported gains. That affects validity and reproducibility, but it does not make the reconstructed output equal to the input by construction, and the correct response is to rerun the experiments with a receiver-computable noise estimator. Section VII also discloses genuine limitations regarding model size, simulated channels, and reward tuning, which is consistent with an empirical systems paper rather than a circular derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claims rely on several unspecified or unstated mechanisms rather than on a derivation. The free parameters include reward-shaping coefficients whose values are never reported, hand-chosen LoRA ranks, and exploration constants. The axioms include the receiver's ability to know the clean transmitted latent for noise-level estimation, a latent-space diffusion training setup that is never described, and a continuous-action DQN outside the standard discrete-action formalism. No new physical entities are introduced; the only invented mechanism is an automatic channel-type detector used for adapter selection, which has no algorithm or evidence.

free parameters (4)
  • RL reward coefficients alpha, beta, gamma, delta = not reported
    Eq. (11) defines the reward as alpha*Delta_L + beta*Delta_Acc + gamma*B_significant + delta*B_entropy; Section IV.A.3 says the scaling factors are 'tuned' but no values or tuning procedure are given. The learned task-balancing behavior depends on them.
  • Module-specific LoRA ranks and scaling factors = encoder 16, decoder 16, EDM 8, classifier 4; alpha_c = alpha_hat_c / r_c
    Chosen by hand in Section VI.A.3 and V.B; adaptation performance depends on these values and no sensitivity analysis is reported.
  • Significance bonus threshold and exploration mixture = 0.05 threshold; 70% uniform / 30% beta mixture; epsilon decays 1.0 to 0.05 over 50,000 steps
    Hand-specified exploration and reward-shaping constants in Section IV.A.3 and IV.B.3; no ablations show their effect on results.
  • Diffusion denoising steps and EDM noise schedule = not specified
    Section III.B references EDM but does not state the number of sampling steps, preconditioning parameters, or the training loss used for the latent denoiser.
assumptions (5)
  • ad hoc to paper The receiver can compute sigma_max = ||z_ch - z_norm||_2 to set the EDM noise level.
    Algorithm 1 line 9 uses the clean transmitted latent z_norm to estimate the noise level, but z_norm is not available at the receiver. This is an unstated and generally false assumption.
  • domain assumption The EDM denoiser is trained on latent codes of the Swin encoder under channel noise and can be integrated end-to-end.
    The paper never specifies the diffusion training loss, the distribution of corrupted latents, or how the denoiser is initialized; Section III.B assumes this works.
  • ad hoc to paper Standard DQN can output continuous actions satisfying lambda_recon + lambda_cls = 1.
    Section IV.B uses a DQN with Softplus output for a continuous action space, which is outside the standard discrete-action DQN formalism; no continuous-action adaptation is described.
  • domain assumption The simulated channels in Eq. (1) capture the relevant dynamics of real wireless environments.
    The paper itself lists this as a limitation: mobility-induced Doppler and large-scale interference are omitted in Section VII.
  • domain assumption Task labels y are available for computing L_cls and the reward during training.
    Algorithm 1 requires ground-truth labels at the receiver; this is standard for supervised training but unstated for deployment.
invented entities (1)
  • Automatic channel-type detection module
    purpose: Selects the right LoRA adapter instantaneously at inference without manual intervention.
    Section V.E claims 'automatic channel detection triggers immediate adapter activation', but no detection algorithm, training data, accuracy, or implementation is given. No independent evidence exists in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TOAST: Task-Oriented Adaptive Semantic Transmission over Dynamic Wireless Environments." pith.science (2026). https://pith.science/paper/VXIK4KZE

@misc{pith2026250621900,
  author       = {Pith},
  title        = {Pith review of: TOAST: Task-Oriented Adaptive Semantic Transmission over Dynamic Wireless Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VXIK4KZE}},
  note         = {Machine review of arXiv:2506.21900}
}
read the original abstract

The evolution toward 6G networks demands a fundamental shift from bit-centric transmission to semantic-aware communication that emphasizes task-relevant information. This work introduces TOAST (Task-Oriented Adaptive Semantic Transmission), a unified framework designed to address the core challenge of multi-task optimization in dynamic wireless environments through three complementary components. First, we formulate adaptive task balancing as a Markov decision process, employing deep reinforcement learning to dynamically adjust the trade-off between image reconstruction fidelity and semantic classification accuracy based on real-time channel conditions. Second, we integrate module-specific Low-Rank Adaptation (LoRA) mechanisms throughout our Swin Transformer-based joint source-channel coding architecture, enabling parameter-efficient fine-tuning that dramatically reduces adaptation overhead while maintaining full performance across diverse channel impairments including Additive White Gaussian Noise (AWGN), fading, phase noise, and impulse interference. Third, we incorporate an Elucidating diffusion model that operates in the latent space to restore features corrupted by channel noises, providing substantial quality improvements compared to baseline approaches. Extensive experiments across multiple datasets demonstrate that TOAST achieves superior performance compared to baseline approaches, with significant improvements in both classification accuracy and reconstruction quality at low Signal-to-Noise Ratio (SNR) conditions while maintaining robust performance across all tested scenarios.

Figures

Figures reproduced from arXiv: 2506.21900 by the authors.

Figure 1
Figure 1. System overview of the proposed semantic communication framework. The input image is encoded via a Swin Transformer to produce a latent code [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Module-Specific Low-Rank Adaptation (LoRA) architecture for the [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Training convergence analysis: PSNR evolution across epochs for [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Training convergence analysis: Classification accuracy evolution [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of reconstruction quality on SVHN dataset at [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: PSNR performance comparison between TOAST and TOAST without [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: SSIM performance comparison between TOAST and TOAST without [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 49 canonical work pages

  1. [7]

    SwinJSCC: Taming Swin Transformer for deep joint source-channel coding,

    K. Yang, S. Wang, J. Dai, X. Qin, K. Niu, and P. Zhang, “SwinJSCC: Taming Swin Transformer for deep joint source-channel coding,” IEEE Trans. Cogn. Commun. Netw. , vol. 11, no. 1, pp. 90–104, Feb. 2025

    work page 2025

  2. [17]

    CDDM: Channel denoising diffusion models for wireless semantic communications,

    T. Wu et al., “CDDM: Channel denoising diffusion models for wireless semantic communications,” IEEE Trans. Wireless Commun. , vol. 23, no. 9, pp. 11 168–11 183, Sept. 2024

    work page 2024

  3. [8]

    Diffusion-aided joint source channel coding for high realism wireless image transmission,

    M. Yang, B. Liu, B. Wang, and H.-S. Kim, “Diffusion-aided joint source channel coding for high realism wireless image transmission,” arXiv preprint arXiv:2404.17736, 2024

    arXiv 2024

  4. [1]

    Beyond transmitting bits: Context, semantics, and task-oriented communications,

    D. G ¨und¨uz et al. , “Beyond transmitting bits: Context, semantics, and task-oriented communications,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 5–41, Jan. 2023

    work page 2023

  5. [2]

    6G networks: Beyond shannon towards semantic and goal-oriented communications,

    E. C. Strinati et al. , “6G networks: Beyond shannon towards semantic and goal-oriented communications,” Comput. Netw., vol. 190, p. 107930, May 2021

    work page 2021

  6. [3]

    Learning task-oriented semantic communi- cation for edge computing,

    H. Seo, Z. Qin, and X. Tao, “Learning task-oriented semantic communi- cation for edge computing,” in Proc. IEEE Wirel. Commun. Netw. Conf. (WCNC), 2023, pp. 1–6

    work page 2023

  7. [4]

    Semantic communications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” arXiv preprint arXiv:2201.01389 , 2021

    arXiv 2021

  8. [5]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, Sept. 2019

    work page 2019

Show all 55 references
  1. [6]

    DeepJSCC-f: Deep joint source-channel coding of images with feedback,

    D. B. Kurka and D. G ¨und¨uz, “DeepJSCC-f: Deep joint source-channel coding of images with feedback,” IEEE J. Sel. Areas Inf. Theory , vol. 1, no. 1, pp. 178–193, May 2020

    2020

  2. [9]

    Vector quantized se- mantic communication system,

    Q. Fu, H. Xie, Z. Qin, G. Slabaugh, and X. Tao, “Vector quantized se- mantic communication system,” IEEE Wireless Commun. Lett. , vol. 12, no. 6, pp. 982–986, Jun. 2023

    2023

  3. [10]

    Deep joint source-channel coding for wireless image transmission with semantic importance,

    Q. Sun, C. Guo, Y . Yang, J. Chen, R. Tang, and C. Liu, “Deep joint source-channel coding for wireless image transmission with semantic importance,” in Proc. IEEE Veh. Technol. Conf. (VTC-Fall) , 2022, pp. 1–7

    2022

  4. [11]

    Feature importance-aware task-oriented semantic trans- mission and optimization,

    Y . Wang et al., “Feature importance-aware task-oriented semantic trans- mission and optimization,” IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 4, pp. 1175–1189, Aug. 2024

    2024

  5. [12]

    Scene graph semantic inference for image and text matching,

    J. Pei, K. Zhong, Z. Yu, L. Wang, and K. Lakshmanna, “Scene graph semantic inference for image and text matching,” ACM Trans. Asian Low-Resour. Lang. Inf. Process. , vol. 22, no. 5, May 2023. [Online]. Available: https://doi.org/10.1145/3563390

    2023 doi

  6. [13]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 35, 2022, pp. 26 565–26 577

    2022

  7. [14]

    Task- oriented communications for 6G: Vision, principles, and technologies,

    Y . Shi, Y . Zhou, D. Wen, Y . Wu, C. Jiang, and K. B. Letaief, “Task- oriented communications for 6G: Vision, principles, and technologies,” IEEE Wireless Commun., vol. 30, no. 3, pp. 78–85, Jun. 2023. Preprint submitted to a journal for possible publication, copyright may be...

    2023

  8. [15]

    Semantic communications for image recovery and classification via deep joint source and channel coding,

    Z. Lyu, G. Zhu, J. Xu, B. Ai, and S. Cui, “Semantic communications for image recovery and classification via deep joint source and channel coding,” IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 8388–8404, Aug. 2024

    2024

  9. [16]

    Task-oriented scene graph-based semantic communications with adaptive channel coding,

    S. Sun, Z. Qin, H. Xie, and X. Tao, “Task-oriented scene graph-based semantic communications with adaptive channel coding,” IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 17 070–17 083, Nov. 2024

    2024

  10. [18]

    A hybrid wireless image transmission scheme with diffusion,

    X. Niu, X. Wang, D. G ¨und¨uz, B. Bai, W. Chen, and G. Zhou, “A hybrid wireless image transmission scheme with diffusion,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC) , 2023, pp. 86–90

    2023

  11. [19]

    A multi-task oriented semantic communication framework for autonomous vehicles,

    E. Eldeeb, M. Shehab, and H. Alves, “A multi-task oriented semantic communication framework for autonomous vehicles,” IEEE Wireless Commun. Lett., vol. 13, no. 12, pp. 3469–3473, Dec. 2024

    2024

  12. [20]

    Task-oriented explainable semantic communications,

    S. Ma et al. , “Task-oriented explainable semantic communications,” IEEE Trans. Wireless Commun. , vol. 22, no. 12, pp. 9248–9262, Dec. 2023

    2023

  13. [21]

    Semantic transfer between different tasks in the semantic communication system,

    Q. Wu, F. Liu, H. Xia, and T. Zhang, “Semantic transfer between different tasks in the semantic communication system,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC) , 2022, pp. 566–571

    2022

  14. [22]

    Joint source-channel coding for channel-adaptive digital semantic communications,

    J. Park, Y . Oh, S. Kim, and Y .-S. Jeon, “Joint source-channel coding for channel-adaptive digital semantic communications,” IEEE Trans. Cogn. Commun. Netw., vol. 11, no. 1, pp. 75–89, Feb. 2025

    2025

  15. [23]

    Channel- aware deep joint source-channel coding for multi-task oriented semantic communication,

    B. Wang, R. Gu, W. Xu, F. Jiang, M. Li, and S. Wang, “Channel- aware deep joint source-channel coding for multi-task oriented semantic communication,” IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1521– 1525, May 2025

    2025

  16. [24]

    Deep reinforced segment selection and equalization for task-oriented semantic communication,

    J. Seon et al., “Deep reinforced segment selection and equalization for task-oriented semantic communication,” IEEE Commun. Lett. , vol. 28, no. 8, pp. 1865–1869, Aug. 2024

    2024

  17. [25]

    Adaptable semantic compression and resource allocation for task-oriented communications,

    C. Liu, C. Guo, Y . Yang, and N. Jiang, “Adaptable semantic compression and resource allocation for task-oriented communications,” IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 3, pp. 769–782, Jun. 2023

    2023

  18. [26]

    Joint task and data-oriented semantic communications: A deep separate source-channel coding scheme,

    J. Huang, D. Li, C. Huang, X. Qin, and W. Zhang, “Joint task and data-oriented semantic communications: A deep separate source-channel coding scheme,” IEEE Internet Things J., vol. 11, no. 2, pp. 2255–2272, Jan. 2024

    2024

  19. [27]

    Rate-adaptive coding mechanism for semantic communications with multi-modal data,

    Y . He, G. Yu, and Y . Cai, “Rate-adaptive coding mechanism for semantic communications with multi-modal data,” IEEE Trans. Commun., vol. 72, no. 3, pp. 1385–1400, Mar. 2023

    2023

  20. [28]

    Digital-analog transmission framework for task-oriented semantic communications,

    Y . Fu, W. Cheng, W. Zhang, and J. Wang, “Digital-analog transmission framework for task-oriented semantic communications,” IEEE Netw. , vol. 38, no. 6, pp. 81–88, Nov. 2024

    2024

  21. [29]

    Learning task-oriented communication for edge inference: An information bottleneck approach,

    J. Shao, Y . Mao, and J. Zhang, “Learning task-oriented communication for edge inference: An information bottleneck approach,” IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 197–211, Jan. 2022

    2022

  22. [30]

    Toward semantic communications: Deep learning-based image semantic coding,

    D. Huang, F. Gao, X. Tao, Q. Du, and J. Lu, “Toward semantic communications: Deep learning-based image semantic coding,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 55–71, Jan. 2023

    2023

  23. [31]

    Robust information bottleneck for task-oriented communication with digital modulation,

    S. Xie, S. Ma, M. Ding, Y . Shi, M. Tang, and Y . Wu, “Robust information bottleneck for task-oriented communication with digital modulation,” IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2577–2591, Aug. 2023

    2023

  24. [32]

    Adaptive modulation and retransmission scheme for semantic communication systems,

    H. Gao, G. Yu, and Y . Cai, “Adaptive modulation and retransmission scheme for semantic communication systems,” IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 1, pp. 150–163, Feb. 2024

    2024

  25. [33]

    WITT: A wireless image transmission transformer for semantic communications,

    K. Yang, S. Wang, J. Dai, K. Tan, K. Niu, and P. Zhang, “WITT: A wireless image transmission transformer for semantic communications,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP) , 2023, pp. 1–5

    2023

  26. [34]

    Channel-adaptive wireless image transmission with OFDM,

    H. Wu, Y . Shao, K. Mikolajczyk, and D. G ¨und¨uz, “Channel-adaptive wireless image transmission with OFDM,” IEEE Wireless Commun. Lett., vol. 11, no. 11, pp. 2400–2404, Nov. 2022

    2022

  27. [35]

    OFDM-guided deep joint source channel coding for wireless multipath fading channels,

    M. Yang, C. Bian, and H.-S. Kim, “OFDM-guided deep joint source channel coding for wireless multipath fading channels,” IEEE Trans. Cogn. Commun. Netw., vol. 8, no. 2, pp. 584–599, Jun. 2022

    2022

  28. [36]

    Deep joint source-channel coding for wireless image transmission with adaptive rate control,

    M. Yang and H.-S. Kim, “Deep joint source-channel coding for wireless image transmission with adaptive rate control,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP) , 2022, pp. 5193–5197

    2022

  29. [37]

    Predictive and adaptive deep coding for wireless image transmission in semantic communication,

    W. Zhang, H. Zhang, H. Ma, H. Shao, N. Wang, and V . C. M. Leung, “Predictive and adaptive deep coding for wireless image transmission in semantic communication,” IEEE Trans. Wireless Commun. , vol. 22, no. 8, pp. 5486–5501, Aug. 2023

    2023

  30. [38]

    Robust semantic communications with masked VQ-V AE enabled codebook,

    Q. Hu, G. Zhang, Z. Qin, Y . Cai, G. Yu, and G. Y . Li, “Robust semantic communications with masked VQ-V AE enabled codebook,”IEEE Trans. Wireless Commun., vol. 22, no. 12, pp. 8707–8722, Dec. 2023

    2023

  31. [39]

    Infomax neural joint source-channel coding via adversarial bit flip,

    Y . Song, M. Xu, L. Yu, H. Zhou, S. Shao, and Y . Yu, “Infomax neural joint source-channel coding via adversarial bit flip,” in Proc. AAAI Conf. Artif. Intell., vol. 34, no. 04, Apr. 2020, pp. 5834–5841

    2020

  32. [40]

    DeepJSCC- Q: Constellation constrained deep joint source-channel coding,

    T.-Y . Tung, D. B. Kurka, M. Jankowski, and D. G ¨und¨uz, “DeepJSCC- Q: Constellation constrained deep joint source-channel coding,” IEEE J. Sel. Areas Inf. Theory , vol. 3, no. 4, pp. 720–731, Dec. 2022

    2022

  33. [41]

    Learning based joint coding- modulation for digital semantic communication systems,

    Y . Bo, Y . Duan, S. Shao, and M. Tao, “Learning based joint coding- modulation for digital semantic communication systems,” in Proc. Int. Conf. Wireless Commun. Signal Process. (WCSP) , 2022, pp. 1–6

    2022

  34. [42]

    Multi-level semantic-aware communication for multi-task image transmission,

    Z. Zhang, Q. Yang, S. He, and J. Chen, “Multi-level semantic-aware communication for multi-task image transmission,” J. Franklin Inst., vol. 362, no. 9, p. 107598, Jun. 2025

    2025

  35. [43]

    Personalized saliency in task-oriented semantic commu- nications: Image transmission and performance analysis,

    J. Kang et al., “Personalized saliency in task-oriented semantic commu- nications: Image transmission and performance analysis,” IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 186–201, Jan. 2023

    2023

  36. [44]

    Task-oriented communications for nextg: End-to-end deep learning and ai security aspects,

    Y . E. Sagduyu, S. Ulukus, and A. Yener, “Task-oriented communications for nextg: End-to-end deep learning and ai security aspects,” IEEE Wireless Commun., vol. 30, no. 3, pp. 52–60, Jun. 2023

    2023

  37. [45]

    Deep learning- enabled semantic communication systems with task-unaware transmitter and dynamic data,

    H. Zhang, S. Shao, M. Tao, X. Bi, and K. B. Letaief, “Deep learning- enabled semantic communication systems with task-unaware transmitter and dynamic data,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 170–185, Jan. 2023

    2023

  38. [46]

    Latency-aware generative semantic communications with pre-trained diffusion models,

    L. Qiao, M. B. Mashhadi, Z. Gao, C. H. Foh, P. Xiao, and M. Bennis, “Latency-aware generative semantic communications with pre-trained diffusion models,” IEEE Wireless Commun. Lett. , vol. 13, no. 10, pp. 2652–2656, Oct. 2024

    2024

  39. [47]

    Latent diffusion model-enabled low-latency semantic communication in the presence of semantic ambiguities and wireless channel noises,

    J. Pei, C. Feng, P. Wang, H. Tabassum, and D. Shi, “Latent diffusion model-enabled low-latency semantic communication in the presence of semantic ambiguities and wireless channel noises,” IEEE Trans. Wireless Commun., vol. 24, no. 5, pp. 4055–4072, May 2025

    2025

  40. [48]

    A multi-task semantic commu- nication system for natural language processing,

    Y . Sheng, F. Li, L. Liang, and S. Jin, “A multi-task semantic commu- nication system for natural language processing,” in Proc. IEEE Veh. Technol. Conf. (VTC-Fall). IEEE, 2022, pp. 1–5

    2022

  41. [49]

    A unified hierarchical semantic knowledge base for multi-task semantic commu- nication,

    L. Wang, W. Wu, F. Zhou, F. Tian, Q. Wu, and W. Saad, “A unified hierarchical semantic knowledge base for multi-task semantic commu- nication,” in Proc. IEEE Int. Conf. Commun. (ICC) . IEEE, 2024, pp. 2937–2943

    2024

  42. [50]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu et al., “LoRA: Low-rank adaptation of large language models,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2022

    2022

  43. [51]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 33, 2020, pp. 6840–6851

    2020

  44. [52]

    Detection and imputation-based two-stage denoising diffusion power system measurement recovery under cyber-physical uncertainties,

    J. Pei, J. Wang, D. Shi, and P. Wang, “Detection and imputation-based two-stage denoising diffusion power system measurement recovery under cyber-physical uncertainties,” IEEE Trans. Smart Grid , vol. 15, no. 6, pp. 5965–5980, Nov. 2024

    2024

  45. [53]

    Neue methoden zur approximativen integration der differentialgleichungen einer unabh ¨angigen ver ¨anderlichen,

    K. Heun et al. , “Neue methoden zur approximativen integration der differentialgleichungen einer unabh ¨angigen ver ¨anderlichen,” Z. Math. Phys, vol. 45, pp. 23–38, 1900

    1900

  46. [54]

    Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV) , 2015, pp. 1026–1034

    2015

  47. [55]

    Understanding the difficulty of training deep feedforward neural networks,

    X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proc. 13th Int. Conf. Artif. Intell. Statist. JMLR, 2010, pp. 249–256

    2010

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.