Pith. sign in

REVIEW 4 major objections 5 minor 39 references

A query-driven image semantic codec that freezes a large vision-language model and steers hybrid beamforming can preserve user-intent content on unseen object categories better than pixel codecs and prior semantic schemes, especially at ver

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

2026-07-31 12:29 UTC pith:Z4KNOROA

load-bearing objection Solid cross-layer systems paper: frozen CLIP + query FiLM + bandit-weighted hybrid BF, with real low-SNR gains on a category-split protocol that is useful but not airtight OOD. the 4 major comments →

arxiv 2607.28276 v1 pith:Z4KNOROA submitted 2026-07-30 cs.IT cs.SYeess.IVeess.SYmath.IT

Generalized Query-Oriented Image Semantic Coding Empowered by Large AI Models and Semantic-Aware Hybrid Beamforming

Sin-Yu Huang , Vincent W.S. Wong This is my paper
classification cs.IT cs.SYeess.IVeess.SYmath.IT
keywords semantic communicationquery-oriented image codinglarge AI modelsCLIPhybrid beamformingMIMO-OFDMcontextual banditzero-shot generalization
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.

Wireless systems usually send every bit of an image with equal care, even when a human viewer only cares about one part of the scene. This paper claims that if the receiver first sends a short text query stating that intent, the transmitter can extract only the matching visual features with a frozen pretrained vision-language model, quantize them, and then deliberately protect the most important of those features with a specially weighted hybrid beamformer in a large MIMO-OFDM link. The reconstructed image is judged not by pixel error alone but by whether an independent vision-language model still answers the same query correctly. On held-out object categories the authors never trained on, the full system keeps those answers accurate farther into the noise floor than JPEG2000, a strong joint source-channel transformer, and an earlier query-plus-diffusion baseline. A sympathetic reader cares because intent is dynamic in surveillance, remote assistance, and AR; a codec that generalizes without fine-tuning and that spends radio resources on what the user actually asked about is closer to usable human-centric 6G.

Core claim

When a frozen CLIP backbone extracts multi-layer image features, FiLM-aligns them to a user text query, segment-wise quantizes the result, and a contextual-bandit weight module drives a weighted-MMSE hybrid beamformer, the end-to-end system reconstructs images whose query-conditioned answers match the original more often than strong baselines on object categories excluded from training, with the largest gains appearing below 0 dB SNR.

What carries the argument

Semantic-aware hybrid beamforming (SA-HBF): a weight module produces a Dirichlet-sampled importance vector over OFDM subcarriers; that vector enters a weighted MMSE objective whose hybrid precoder/combiner solutions preferentially protect the query-relevant feature blocks.

Load-bearing premise

Before the detailed image is sent, the user can already supply an accurate text query from general scene knowledge, and that short query arrives error-free on a separate control channel.

What would settle it

Retrain and re-evaluate the identical architecture on the same VQA animal/human hold-out splits but replace the true user query with a randomly chosen unrelated query (or drop the query entirely); if the answer-match-rate advantage over the no-alignment and uniform-weight ablations disappears at SNR = -25 dB, the central intent-preservation claim fails.

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

If this is right

  • Large-scale MIMO-OFDM links can treat semantic importance weights as a cross-layer interface instead of only maximizing sum-rate.
  • Freezing a pretrained vision-language backbone inside the codec removes the need to fine-tune on every new object category.
  • At low SNR the encoder is forced to spend its limited capacity on query-relevant patches, automatically trading global fidelity for intent fidelity.
  • The same three-stage training (semantic codec, then bandit beamformer, then joint fine-tune) can be reused for other query-conditioned modalities.
  • Multi-user extensions become natural once each user’s query defines its own importance map over the shared spatial-frequency resources.

Where Pith is reading between the lines

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

  • If the control-channel assumption is relaxed, a lightweight joint source-channel code for the query itself would become the next bottleneck and could erase the reported low-SNR gains.
  • The same FiLM-plus-bandit pattern could be dropped onto video or point-cloud streams where temporal consistency of the query-relevant object matters more than single-frame pixel error.
  • Because the weight module is trained with negative total loss as reward, any future perceptual or task loss can be swapped in without redesigning the beamformer.
  • Deployment cost is dominated by the per-image WMMSE solve; a learned warm-start or codebook of hybrid matrices conditioned on the weight vector would be a practical next engineering step.

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

4 major / 5 minor

Summary. The paper proposes a query-oriented image semantic coding (QO-ISC) system for large-scale MIMO-OFDM. A frozen CLIP backbone extracts multi-layer image and text features; FiLM aligns them to a user text query; segment-wise vector quantization compresses the aligned embeddings; and a decoder reconstructs the image. Semantic importance is scored by a weight module trained as a Dirichlet contextual bandit, and the resulting per-subcarrier weights drive a weighted-MMSE hybrid beamformer (SA-HBF). Training is three-stage (codec/codebook, then SA-HBF, then joint fine-tuning) with an LLaVA-based user-intent relevance loss plus ℓ1, adversarial, and quantization terms. On VQA category hold-outs (train non-animal/test animal; train non-human/test human), the system reports higher BLIP-2 answer-match rate and lower BLIP-2 user-intent loss than JPEG2000, SwinJSCC, a query-guided diffusion baseline, and ablations without query alignment or importance weighting, with the largest gains at very low SNR.

Significance. If the reported gains hold under a cleaner OOD protocol, the work is a solid systems contribution at the intersection of semantic coding and hybrid MIMO-OFDM: it jointly addresses intent-conditioned feature selection, frozen-LAM generalization, and semantic-aware HBF, with useful ablations (uniform/random weights, no query alignment, power-allocation variant), an evaluator (BLIP-2) not used in training, and explicit complexity accounting (Table I, online runtimes). The cross-layer use of learned semantic weights inside WMMSE-HBF is a concrete and reusable design idea. The main significance risk is that the headline “unseen object category” claim may partly reflect residual scene/query leakage in the VQA split rather than true category generalization; resolving that would substantially strengthen the paper for a top venue.

major comments (4)
  1. [§VI.A, Figs. 7a–d] The central generalization claim (abstract, §I Q3, §VI.A, Figs. 7a–d) rests on training on non-animal (resp. non-human) VQA subsets and testing only on animal (resp. human) samples. VQA scenes are multi-object; excluding “animal-related samples” does not preclude co-occurring train-category objects, shared backgrounds, or repeated question templates that a frozen CLIP/LLaVA stack can exploit. The paper reports neither co-occurrence / leakage statistics nor an external OOD set (e.g., animals from a corpus disjoint from VQA). Without that, the large margins vs. SwinJSCC (trained from scratch on the same split) and vs. the fine-tuned query baseline can overstate true out-of-category generalization. Please either (i) quantify residual leakage and add a stricter isolation or external OOD evaluation, or (ii) reframe the claim as category-holdout within VQA rather than strong zero-shot generali
  2. [§VI, Figs. 7–9, 11–12; abstract] Headline numerical gains (e.g., +4.8% vs. uniform SA-HBF, +9% vs. no query alignment, +8% vs. [20], +30% vs. SwinJSCC at SNR = −25 dB) are reported as point estimates averaged over channel realizations, with no error bars, multi-seed training variance, or statistical tests on the match-rate / L_BLIP-2_user curves. At the operating points where the paper’s contribution is most visible (very low SNR), small training or split variability could move the ranking among close ablations. Please add multi-seed (or at least multi-run) statistics and uncertainty bands on Figs. 7–9 and 11–12 for the metrics that support the abstract claims.
  3. [§II (incl. footnote 1); Eqs. (3)–(4)] The system model assumes the user issues a correct text query from “general knowledge of the scene” before transmission, and that the query is delivered reliably on a separate control channel (§II, footnote 1). FiLM alignment (Eqs. 3–4) and SA-HBF priority both target that query; a wrong, underspecified, or unavailable query would steer capacity toward the wrong semantics. This assumption is load-bearing for the intent-preserving gains but is not stress-tested (mismatched query, empty query, or noisy query). A short sensitivity study—or an explicit limitation with degraded-mode behavior (e.g., fall back to uniform weights / no FiLM)—is needed so readers know when the claimed gains apply.
  4. [§VI.G; Algorithm 4; Fig. 9] Online cost: the paper states WMMSE-HBF takes ~3 s per image on CPU because weights depend on the current features and beamformers are recomputed per image–query pair (§VI.G). For a communication system paper whose contribution includes SA-HBF in large-scale MIMO-OFDM, this is a practical bottleneck relative to the encoder/decoder (ms-scale on GPU). Please discuss amortization (e.g., update beamformers on a slower timescale, codebook of beams, or approximate closed-form updates) or show that a cheaper HBF surrogate retains most of the semantic gain in Fig. 9; otherwise the SA-HBF contribution is mainly offline/simulation-facing.
minor comments (5)
  1. [§II] Notation: both the analog combiner and the quantization codebook family use overlapping calligraphic/roman forms in places; a short notation table or consistent font for C (codebooks) vs. channel-related symbols would help.
  2. [Fig. 6] Fig. 6 is persuasive but only one qualitative example at one SNR; adding one failure case (mismatched query or non-query region) would balance the narrative.
  3. [§V.A, Eqs. (28)–(30), Algorithm 3] The Dirichlet–Thompson / REINFORCE weight module is described clearly, but the text sometimes calls it “Thompson sampling” while the update is a single-sample policy gradient (Eq. 30). Align the naming with what is implemented.
  4. [Table I] Table I GFLOPs and parameter counts are useful; briefly state the batch size / precision and whether CLIP forward passes are included so numbers are reproducible.
  5. [Abstract; References] A few copy-edits: “F urthermore”, broken line wraps in the abstract/intro, and “Prof. of Int’l Conf.” in [25] should be “Proc.”

Circularity Check

0 steps flagged

No significant circularity: empirical system gains are measured with an independent evaluator (BLIP-2) distinct from the training loss (LLaVA), and no claim reduces to its inputs by construction.

full rationale

This is an engineering systems paper whose central claims are empirical simulation comparisons (answer match rate and L_BLIP-2_user on held-out VQA category splits vs. JPEG2000, SwinJSCC, a query-diffusion baseline, and ablations). Training minimizes L_total = L1 + λ_user L_LLaVA_user + λ_gen L_gen + λ_quant L_quant (§III, Eq. 22); evaluation deliberately uses a different frozen model (BLIP-2) and exact-match/BERTScore on answers, with the paper stating this is to avoid self-scoring bias (§VI). The SA-HBF contextual bandit uses reward r = −L_total (Algorithm 3), which is ordinary policy-gradient optimization of the training objective, not a definition that forces the reported BLIP-2 metrics. Hybrid beamforming extends a standard WMMSE/MO procedure [37] with learned subcarrier weights; the weights are not fitted to the test metric. Self-citations ([1] partial conference version; [18] prior text work) are not load-bearing uniqueness or uniqueness-import steps. No equation equates a claimed prediction to a fitted input by construction. Protocol concerns about residual VQA co-occurrence leakage affect generalization validity, not circularity of the derivation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 3 invented entities

The central performance claim rests on standard wireless and ML modeling choices, several hand-set training weights, the assumption that a pre-image text query is available and reliable, and proxy evaluation via frozen VLMs—not on new physical entities. Free parameters are the usual deep-learning knobs; invented ‘entities’ are architectural modules, not ontological additions.

free parameters (4)
  • λ_user, λ_gen, λ_quant, γ = 1.5, 0.5, 0.5, 1.2
    Loss weights and VQ commitment coefficient set to 1.5, 0.5, 0.5, 1.2; directly steer the intent-vs-fidelity tradeoff shown in Figs. 10–11.
  • Selected CLIP layers MS and segment length NL = MS={3,6,9,11}, NL=64, N_cw=64
    MS={3,6,9,11}, NL=64 (L=8), N_cw=64 chosen by design; change transmitted symbol count and multi-scale content.
  • MIMO-OFDM / HBF dimensions and channel geometry = Nt=Nr=64, NRF=2, K=64, NS=2
    Nt=Nr=64, NRF=2, K=64, NS=2, NC=5, NR=10 Laplacian angles—simulation regime defining reported SNR curves.
  • Stage epoch counts, learning rates, bandit rounds T, NHBF iterations
    Training schedule and WMMSE iteration budget affect convergence of codec and weights; values named as inputs to Algs. 1–4 but not fully numeric in text for all.
axioms (7)
  • domain assumption Text query is available at the transmitter from user general scene knowledge and is delivered reliably on a control channel before image features.
    Stated in §II and footnote 1; without it query alignment and SA-HBF targets are undefined.
  • domain assumption Frozen pretrained CLIP (and LLaVA/BLIP-2 evaluators) provide transferable multimodal embeddings for unseen categories.
    Core of Q3 and zero-shot protocol; backbone never fine-tuned (§II, §VI).
  • domain assumption Geometry-based clustered multipath MIMO-OFDM channel with complex Gaussian path gains and Laplacian AoA/AoD.
    Eq. (13)–(15); all SNR and antenna results assume this generative model.
  • domain assumption Analog precoder/combiner entries have constant modulus (phase-shifter HBF); per-subcarrier power constraint ||V[k]||_F^2 ≤ P.
    Constraints (23b)–(23d); shapes SA-HBF design space vs relaxed PA baseline.
  • ad hoc to paper Answer match via BLIP-2 (and BERTScore on open answers) is a valid proxy for preservation of user-intent semantics.
    §VI evaluation design; central quantitative claims are in this proxy, not human ratings.
  • standard math Standard alternating WMMSE + manifold optimization converges to useful hybrid beamformers for weighted MSE (extension of [37]).
    Appendix Algorithm 4; no global optimality proof for nonconvex HBF.
  • ad hoc to paper Dirichlet-policy contextual bandit with Thompson-style sampling and REINFORCE-style updates learns meaningful subcarrier importance.
    §V-A; reward is −L_total under frozen Stage-I codec.
invented entities (3)
  • QO-ISC encoder–decoder with segment-wise codebook and FiLM query alignment on multi-layer CLIP features no independent evidence
    purpose: Extract and quantize query-relevant image embeddings while keeping CLIP frozen for generalization.
    Architectural composition; not a new physical object. Behavior only evidenced inside this paper’s sims.
  • SA-HBF weight module (contextual bandit → per-subcarrier weights → weighted MMSE HBF) no independent evidence
    purpose: Cross-layer bridge so semantically important features get preferential hybrid beamforming.
    Named algorithm unique to this work; independent evidence would need external reimplementation or over-the-air test—neither provided.
  • User-intent relevance loss L_user via frozen LLaVA multimodal hidden states no independent evidence
    purpose: Train reconstruction to match original image semantics under the same text query without using ground-truth answers.
    Eq. (20); training objective invented for this pipeline; eval switches to BLIP-2 analog.

pith-pipeline@v1.2.0-daily-grok45 · 31594 in / 4169 out tokens · 85416 ms · 2026-07-31T12:29:45.542386+00:00 · methodology

0 comments
read the original abstract

Semantic communication is an emerging paradigm that can preserve the meaning of data during transmission. However, human users are often interested in specific semantic content based on their intent, and users' intent is often not considered in current semantic coding design. Moreover, most of the existing semantic models are fine-tuned using specific datasets, which limits their generalization capability. Furthermore, how to prioritize semantically important features in large-scale multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems remains largely unexplored. To address the aforementioned challenges, in this paper, we propose a generalized query-oriented image semantic coding (QO-ISC) framework. In the proposed framework, the transmitter extracts features which are relevant to the user's query and the receiver reconstructs an image based on those features. We use a pretrained large artificial intelligence (AI) model (LAM) to enhance general feature representations. We develop a semantic-aware hybrid beamforming (SA-HBF) algorithm to prioritize semantically important features for large-scale MIMO-OFDM system. When evaluated on unseen object categories within the dataset, simulation results show that our proposed generalized QO-ISC framework achieves better performance than the traditional codec and two state-of-the-art semantic coding schemes.

Figures

Figures reproduced from arXiv: 2607.28276 by Sin-Yu Huang, Vincent W.S. Wong.

Figure 1
Figure 1. Figure 1: An example image to show that the importance of semant [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The system model of the proposed generalized QO-ISC fr [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The weighting process, segment-wise quantization p [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The user-intent relevance loss derived from pretrai [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The contextual bandit framework for semantic-aware [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the images reconstructed by our pro [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Symbol error rate versus the SNR. -25 -20 -15 -10 SNR (dB) 0.1 0.15 0.2 0.25 0.3 0.35 0.4 (a) 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.12 0.14 0.16 0.18 0.2 0.22 (b) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: The user-intent relevance loss (BLIP-2) L BLIP-2 user versus the tunable parameter λuser. between the user-intent relevance loss and other objectives, such as the ℓ1 loss and the adversarial loss. Specifically, the user-intent relevance loss focuses on preserving local semantic details aligned with the query, whereas other loss functions such as ℓ1 and adversarial losses aim to maintain the global structu… view at source ↗
Figure 10
Figure 10. Figure 10: (a) Visualization of the original image. Visualiza [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: The answer match rate versus different types of ques [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: (a) SSIM and (b) FID versus the SNR. Higher SSIM indic [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] 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

39 extracted references

  1. [1]

    Generalized user-oriente d image semantic coding empowered by large vision-language model

    S.-Y . Huang and V . W.S. Wong, “Generalized user-oriente d image semantic coding empowered by large vision-language model. ” in Proc. of IEEE Global Communications Conference (GLOBECOM) , Taipei, Taiwan, Dec. 2025

    2025

  2. [2]

    Semantic communic ation: A survey on research landscape, challenges, and future direc tions,

    T. M. Getu, G. Kaddoum, and M. Bennis, “Semantic communic ation: A survey on research landscape, challenges, and future direc tions,” Proc. of the IEEE , vol. 112, no. 11, pp. 1649–1685, Nov. 2024

    2024

  3. [3]

    Beyond transmitting bits: Conte xt, seman- tics, and task-oriented communications,

    D. G¨ und¨ uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Y ang, A. Y ener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Conte xt, seman- tics, and task-oriented communications,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 5–41, Jan. 2023

    2023

  4. [4]

    Deep learning en abled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning en abled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663–2675, Apr. 2021

    2021

  5. [5]

    BERTScore: Evaluating text generation with BERT,

    T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi , “BERTScore: Evaluating text generation with BERT,” in Proc. of Int’l Conf. on Learning Representations (ICLR) , May 2020

    2020

  6. [6]

    BERT: P re-training of deep bidirectional transformers for language understan ding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: P re-training of deep bidirectional transformers for language understan ding,” in Proc. of North American Chapter of the Association for Computatio nal Linguistics (NAACL-HLT), Minneapolis, MN, Jun. 2019

    2019

  7. [7]

    Semanti c MIMO systems for speech-to-text transmission,

    Z. Weng, Z. Qin, H. Xie, X. Tao, and K. B. Letaief, “Semanti c MIMO systems for speech-to-text transmission,” IEEE Trans. Wireless Commun., vol. 23, no. 12, pp. 18697–18710, Dec. 2024

    2024

  8. [8]

    Ge n- erative semantic communications with foundation models: P erception- error analysis and semantic-aware power allocation,

    C. Xu, M. B. Mashhadi, Y . Ma, R. Tafazolli, and J. Wang, “Ge n- erative semantic communications with foundation models: P erception- error analysis and semantic-aware power allocation,” IEEE J. Sel. Areas Commun., vol. 43, no. 7, pp. 2493–2505, Jul. 2025

    2025

  9. [9]

    A robust i mage semantic communication system with multi-scale vision tra nsformer,

    X. Peng, Z. Qin, X. Tao, J. Lu, and K. B. Letaief, “A robust i mage semantic communication system with multi-scale vision tra nsformer,” IEEE J. Sel. Areas Commun. , vol. 43, no. 4, pp. 1278–1291, Apr. 2025

    2025

  10. [10]

    Toward semanti c communications: Deep learning-based image semantic codin g,

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

    2023

  11. [11]

    Importance-awa re semantic communication in MIMO-OFDM systems using vision transform er,

    J. Park, Y . Oh, J. Park, and Y .-S. Jeon, “Importance-awa re semantic communication in MIMO-OFDM systems using vision transform er,” IEEE Trans. Wireless Commun. , vol. 25, pp. 13494–13510, Mar. 2026

    2026

  12. [12]

    OFDM-bas ed digital semantic communication with importance awareness ,

    C. Liu, C. Guo, Y . Y ang, W. Ni, and T. Q. S. Quek, “OFDM-bas ed digital semantic communication with importance awareness ,” IEEE Trans. Commun., vol. 72, no. 10, pp. 6301–6315, Oct. 2024

    2024

  13. [13]

    A tutorial on extremely large-scal e MIMO for 6G: Fundamentals, signal processing, and applications,

    Z. Wang, J. Zhang, H. Du, D. Niyato, S. Cui, B. Ai, M. Debba h, K. B. Letaief, and H. V . Poor, “A tutorial on extremely large-scal e MIMO for 6G: Fundamentals, signal processing, and applications,” IEEE Commun. Surveys & Tuts. , vol. 26, no. 3, pp. 1560–1605, third quarter 2024

    2024

  14. [14]

    Hybrid beamforming for massive MIMO: A surve y,

    A. F. Molisch, V . V . Ratnam, S. Han, Z. Li, S. L. H. Nguyen, L. Li, and K. Haneda, “Hybrid beamforming for massive MIMO: A surve y,” IEEE Commun. Mag. , vol. 55, no. 9, pp. 134–141, Sep. 2017

    2017

  15. [15]

    Hybrid analog and digital beamfor ming for mmWave OFDM large-scale antenna arrays,

    F. Sohrabi and W. Y u, “Hybrid analog and digital beamfor ming for mmWave OFDM large-scale antenna arrays,” IEEE J. Sel. Areas Com- mun., vol. 35, no. 7, pp. 1432–1443, Jul. 2017

    2017

  16. [16]

    Deep jo int source- channel coding for wireless image transmission,

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

    2019

  17. [17]

    Sema ntic importance-aware communications with semantic correctio n using large language models,

    S. Guo, Y . Wang, J. Y e, A. Zhang, P . Zhang, and K. Xu, “Sema ntic importance-aware communications with semantic correctio n using large language models,” IEEE Trans. Mach. Learning Commun. Netw. , vol. 3, pp. 232–245, Jan. 2025

    2025

  18. [18]

    Leveraging MoE- based large language model for zero-shot multi-task semantic communic ation,

    S.-Y . Huang, R. Liao, and V . W.S. Wong, “Leveraging MoE- based large language model for zero-shot multi-task semantic communic ation,” in Proc. of IEEE Int. Conf. Commun. (ICC) , Montreal, Canada, Jun. 2025

    2025

  19. [19]

    V isual language model-based cross-modal semantic communication systems,

    F. Jiang, C. Tang, L. Dong, K. Wang, K. Y ang, and C. Pan, “V isual language model-based cross-modal semantic communication systems,” IEEE Trans. Wireless Commun. , vol. 24, no. 5, pp. 3937–3948, May 2025

    2025

  20. [20]

    Semantic communication based on large language model for u nderwa- ter image transmission,

    W. Chen, W. Xu, H. Chen, X. Zhang, Z. Qin, Y . Zhang, and Z. H an, “Semantic communication based on large language model for u nderwa- ter image transmission,” IEEE Trans. Mobile Comput. , vol. 25, no. 2, pp. 2060–2075, Feb. 2026

    2060

  21. [21]

    BLIP: Bootstrapping l anguage-image pre-training for unified vision-language understanding an d generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “BLIP: Bootstrapping l anguage-image pre-training for unified vision-language understanding an d generation,” in Proc. of Int’l Conf. on Machine Learning (ICML) , Baltimore, MD, Jul. 2022

    2022

  22. [22]

    Learning transferable visual models from natural language supervision,

    A. Radford et al. , “Learning transferable visual models from natural language supervision,” in Proc. of Int’l Conf. on Machine Learning (ICML), Jul. 2021

    2021

  23. [23]

    Visual instruction tu ning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tu ning,” in Proc. of Adv. Neural Inf. Process. Syst. (NeurIPS) , New Orleans, LA, Dec. 2023

    2023

  24. [24]

    Generative adversarial nets,

    I. J. Goodfellow et al. , “Generative adversarial nets,” in Proc. of Adv. Neural Inf. Process. Syst. (NIPS) , Montreal, Canada, Dec. 2014

    2014

  25. [25]

    VQA: Visual question answering,

    S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Z itnick, and D. Parikh, “VQA: Visual question answering,” in Prof. of Int’l Conf. Comput. Vis. (ICCV) , Santiago, Chile, Dec. 2015. 17

    2015

  26. [26]

    Swi nJSCC: Taming Swin transformer for deep joint source-channel codi ng,

    K. Y ang, S. Wang, J. Dai, X. Qin, K. Niu, and P . Zhang, “Swi nJSCC: Taming Swin transformer for deep joint source-channel codi ng,” IEEE Trans. Cognitive Commun. and Netw. , vol. 11, no. 1, pp. 90–104, Feb. 2025

    2025

  27. [27]

    Multiple xing and channel coding

    3rd Generation Partnership Project (3GPP), “Multiple xing and channel coding.” TS 36.212, v19.3.0, Mar. 2026

    2026

  28. [28]

    U-Net: Convol utional net- works for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-Net: Convol utional net- works for biomedical image segmentation,” in Proc. of Medical Image Computing and Computer-Assisted Intervention (MICCAI) , Munich, Germany, Oct. 2015

    2015

  29. [29]

    Feature-wise transformations,

    V . Dumoulin, E. Perez, N. Schucher, F. Strub, H. d. Vries , A. Courville, and Y . Bengio, “Feature-wise transformations,” Distill, Jul. 2018

    2018

  30. [30]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and k. Kavukcuoglu, “Neural discrete representation learning,” in Proc. of Adv. Neural Inf. Process. Syst., Long Beach, CA, Dec. 2017

    2017

  31. [31]

    Beamforming design for semantic-bit coexisting communic ation sys- tem,

    M. Zhang, G. Zhu, R. Jin, X. Chen, Q. Shi, C. Zhong, and K. H uang, “Beamforming design for semantic-bit coexisting communic ation sys- tem,” IEEE J. Sel. Areas Commun. , vol. 43, no. 4, pp. 1262–1277, Apr. 2025

    2025

  32. [32]

    Lattimore and C

    T. Lattimore and C. Szepesv´ ari, Bandit Algorithms. Cambridge Univer- sity Press, 2020

    2020

  33. [33]

    Thompson sampling for context ual bandits with linear payoffs,

    S. Agrawal and N. Goyal, “Thompson sampling for context ual bandits with linear payoffs,” in Proc. of Int’l Conf. on Machine Learning (ICML), Atlanta, GA, Jun. 2013

    2013

  34. [34]

    S. Kotz, N. Balakrishnan, and N. L. Johnson, Continuous Multivariate Distributions: Models and Applications . Wiley, 2000

    2000

  35. [35]

    Deep Bayesian ban dits show- down: An empirical comparison of Bayesian deep networks for Thomp- son sampling,

    C. Riquelme, G. Tucker, and J. Snoek, “Deep Bayesian ban dits show- down: An empirical comparison of Bayesian deep networks for Thomp- son sampling,” in Proc. of Int’l Conf. on Learning Representations (ICLR), V ancouver, Canada, May 2018

    2018

  36. [36]

    Simple statistical gradient-followi ng algorithms for con- nectionist reinforcement learning,

    R. J. Williams, “Simple statistical gradient-followi ng algorithms for con- nectionist reinforcement learning,” Machine Learning , vol. 8, pp. 229– 256, May 1992

    1992

  37. [37]

    Hybr id beam- forming for millimeter wave systems using the MMSE criterio n,

    T. Lin, J. Cong, Y . Zhu, J. Zhang, and K. B. Letaief, “Hybr id beam- forming for millimeter wave systems using the MMSE criterio n,” IEEE Trans. Commun., vol. 67, no. 5, pp. 3693–3708, May 2019

    2019

  38. [38]

    Information Technology - JPEG 2000 I mage Coding System Part 1: Core Coding System,

    ISO/IEC 15444-1, “Information Technology - JPEG 2000 I mage Coding System Part 1: Core Coding System,” 5th edition, Nov. 2024

    2000

  39. [39]

    The c urious case of neural text degeneration,

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y . Choi, “The c urious case of neural text degeneration,” in Proc. of Int’l Conf. on Learning Representations (ICLR) , Virtual, Apr. 2020. Sin-Yu Huang (Graduate Student Member, IEEE) received the B.S. degree in electrical engineering and the M.S. degree in computer and communication engineering from National C...

    2020