Token Communication in the Era of Large Models: An Information Bottleneck-Based Approach

Dusit Niyato; Hao Wei; Ping Zhang; Wanli Ni; Wenjun Xu; Wen Wang

REVIEW 3 major objections 5 minor 1 cited by

Token Communication in the Era of Large Models: An Information Bottleneck-Based Approach

T0 review · 3 major / 5 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 The paper proposes UniToCom, a unified token communication paradigm in which text, image, and audio are each compressed into latent tokens, transmitted over wireless channels, and decoded by a single multimodal large language model at the…

desk verdict The system idea is coherent, but the GenIB derivation reverses the data processing inequality, so the core theory doesn't support the proposal. read the letter →

arxiv 2507.01728 v1 pith:KXIDSPGX submitted 2025-07-02 eess.SP cs.LG

Hao Wei , Wanli Ni , Wen Wang , Wenjun Xu , Dusit Niyato , Ping Zhang This is my paper

classification eess.SPcs.LG

keywords tokencommunicationgenerativeinformationbottleneckmultimodallargelanguagemodelsemanticnext-tokenpredictionlatentdiffusionjointsource-channelcodingwirelesstransmission

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 is trying to establish that tokens, rather than raw bits or high-dimensional semantic features, can serve as the fundamental units of wireless transmission, enabling one unified system to handle multimodal understanding and generation. It proposes UniToCom, which tokenizes each input modality, transmits the latent tokens over fading channels, and lets a causal Transformer-based multimodal large language model (MLLM) at the receiver interpret and generate all modalities through next-token prediction and next-token diffusion. To learn good tokens, it introduces the generative information bottleneck (GenIB) principle, a rate-distortion-style objective that compresses the source while preserving its ability to be reconstructed, along with a variant called $\sigma$-GenIB that stabilizes autoregressive training by coupling deterministic and stochastic latent representations. If the claims hold, a single token-oriented transceiver could replace separate codecs for text, image, and audio and degrade gracefully under dynamic channel conditions. A careful reader would care because the proposal directly connects the token-based internals of large generative models with the physical layer of wireless communication.

What carries the argument

The central object is the GenIB principle, a generative information bottleneck that replaces the task-label-dependent information bottleneck with a rate-distortion tradeoff aimed at source restoration and generation: minimize $\xi I(X;T) - I(\hat{T};X)$, where $T$ is the transmitted token and $\hat{T}$ is the received token. Variational bounds convert this into a tractable loss: the rate term is bounded by $D_{\mathrm{KL}}(p_\alpha(t|x) \| \mathcal{N}(t|0,I))$, and the distortion term is replaced by a variational reconstruction likelihood $\int p(t,x)\log q_\beta(x|t)\,dt\,dx$, computed with mean-square-error or cross-entropy losses. The $\sigma$-GenIB variant parameterizes the stochastic latent as $t = \mu + \sigma \odot \epsilon$ with fixed $\sigma$ and a hyperparameter $C_\sigma$, and mixes a deterministic reconstruction $F_\beta(\mu)$ with the stochastic reconstruction $F_\beta(t)$ under a weighting $\lambda$. At the receiver, a causal Transformer-based latent MLLM unifies discrete tokens through a softmax language-modeling head and continuous tokens through a diffusion head, all under the next-token prediction paradigm.

What would settle it

For a simple Markov chain $X \to T \to \hat{T}$, compute $I(X;T)$ and $I(\hat{T};X)$ empirically, for example with a binary source, a noisy channel from $X$ to $T$, and any deterministic mapping from $T$ to $\hat{T}$; the second mutual information will be strictly smaller, directly contradicting the inequality in Eq. (9b) on which the GenIB lower bound depends.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that UniToCom enables scalable and generalizable communication by making tokens the transmission units: each source $x_m$ is mapped by a tokenizer $F_m$ into latent tokens $t_m$, jointly channel-coded and modulated into $z_m$, sent over a Rayleigh fading channel, and then demodulated and decoded into received tokens $\hat{t}_m$, which a single latent MLLM processes to produce outputs across text, image, and audio. The tokenizer is trained with the GenIB objective, which minimizes the rate $I(X;T)$ subject to a lower bound on the generative informativeness $I(\hat{T};X)$, and the paper derives a variational loss with a KL regularization term plus a reconstruction term. The $\sigma$-GenIB variant fixes the variance of the stochastic latent and adds a deterministic-path reconstruction term to prevent variance collapse. Simulation results on CLEVR for visual question answering, MS-COCO for text-to-image generation, and LibriSpeech for speech recognition show that UniToCom outperforms vanilla semantic-communication baselines and traditional separate codecs, and that a moderate token length (32 tokens for images) yields the best balance of accuracy, communication efficiency, and computational cost.

Load-bearing premise

The derivation of the GenIB loss relies on the claim that the received token contains at least as much information about the source as the transmitted token itself did, but information theory says the received token can only contain less, so the bound that produces the training objective is not valid as written.

Editorial extensions

If this is right

A single UniToCom transceiver could serve multiple modalities with one MLLM-based decoder, eliminating the need for separate source and channel codecs for text, image, and audio.
The reported token-length results imply that moderate compression (32 image tokens rather than 128) improves both task accuracy and computational efficiency, suggesting a sweet spot in the rate-distortion tradeoff.
Under high SNR, UniToCom approaches the performance upper bound, indicating that the learned token representations become the limiting factor only in favorable channels.
The unified next-token paradigm means the same receiver architecture can perform both understanding (e.g., VQA, ASR) and generation (e.g., text-to-image) tasks without task-specific modules.
If the GenIB objective trains tokenizers that preserve generative informativeness, the approach could extend naturally to other modalities such as video by adding another tokenizer and diffusion head.

Reading between the lines

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

The empirical gains attributed to GenIB may not require the information-bottleneck derivation as written, since the architecture (a VAE-style tokenizer plus a pretrained MLLM) could account for much of the improvement; a corrected theoretical bound would be needed to isolate the IB contribution.
The optimal token length likely depends on channel conditions and task complexity, so a practical system might adapt the number of tokens per source symbol to the instantaneous SNR, extending the adaptive coding and modulation module.
The paradigm suggests a token-level joint source-channel coding view in which the MLLM's prior knowledge acts as a form of implicit error correction, a property that could be tested by ablating the MLLM backbone or by sending random token sequences.
Because tokens are modality-agnostic once learned, the same transmitter-receiver pair could be reused for new modalities without retraining the MLLM, provided a tokenizer for the new modality is trained with the GenIB objective.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version: the paper has a coherent system idea—unified token communication across text, image, and audio with a Transformer-based MLLM receiver—but the central GenIB derivation reverses the data processing inequality, so the theory as written doesn't support the proposal. The system-level novelty is real, but the theoretical contribution doesn't hold up.

What's genuinely new: treating tokens as the transmission unit across three modalities with a single latent MLLM receiver, following the Transfusion architecture. That's a sensible direction and the paper's architecture is clearly described. The simulation tasks (CLEVR VQA, MS-COCO text-to-image, LibriSpeech ASR) are standard and the comparisons to JSCC baselines are appropriate.

The load-bearing flaw is Eq. (9b). Along the Markov chain X → T → T̂, the data processing inequality gives I(T̂;X) ≤ I(X;T). The paper asserts the reverse, and the right-hand side of (9b) is exactly I(X;T). This invalidates the variational lower bound and the GenIB loss in (12); the final objective reduces to a β-VAE-style reconstruction-plus-KL loss, which is known. So the 'GenIB principle' is not derived from the stated constrained optimization.

The second issue: Eq. (13) computes the distortion on pre-channel tokens t and μ, while Algorithm 1 passes t̂ through the channel and reconstructs from it. The loss never includes the channel output, so the channel noise doesn't enter the gradient. That's a real mismatch between the algorithm and the equations.

The empirical section is not independently checkable: no code, no numeric tables, and the figures are hard to read. That doesn't help.

Who should read this: people working on JSCC and semantic communications who want a unified token interface. They'll get a plausible architecture and a clear system narrative. They should not rely on the GenIB derivation, and they should be cautious about the training loss claim.

My recommendation: don't accept in current form. The mathematical error is precise and identifiable, so the paper deserves a serious referee—an editor should send it out rather than desk reject—but the review should come back with a clear reject unless the theoretical derivation is rebuilt and the training loop is corrected to actually include the channel. If you're looking for a cautionary example of a reversed inequality, it's useful; as a citable contribution, no.

Referee Report

3 major / 5 minor

Summary. The paper proposes UniToCom, a unified token-based communication paradigm in which text, image, and audio sources are tokenized, transmitted over wireless channels, and processed by a causal Transformer-based multimodal large language model (MLLM) at the receiver. The authors introduce a 'generative information bottleneck' (GenIB) principle for tokenization, together with a variant called σ-GenIB that mixes deterministic and stochastic latent representations to counter variance collapse. Simulations on CLEVR-VQA, MS-COCO text-to-image generation, and LibriSpeech-ASR claim that UniToCom outperforms conventional SemCom schemes under Rayleigh fading. The central theoretical derivation is in Section III, where the GenIB objective is converted into a variational loss.

Significance. If the theoretical derivation were sound, GenIB-based tokenization for token communication with MLLMs could be a valuable step toward unified multimodal semantic communication. The paper addresses a timely problem, and the system architecture is clearly presented. However, the core derivation contains a fundamental error: Eq. (9b) reverses the data processing inequality, so the variational objective does not follow from the stated constrained optimization. Moreover, the training procedure in Algorithm 1 appears to compute the loss on pre-channel tokens, so channel effects may not enter the gradient. These issues undermine the paper's main theoretical and practical claims as written. The empirical validation is also difficult to assess because no numerical results are provided and the figures are not legible in the submitted text.

major comments (3)

[Section III-B, Eq. (9b)] The inequality I(T̂;X) ≥ ∫ p(t,x) log p(x|t) dt dx + H(X) is asserted to hold due to the data processing inequality, but the right-hand side equals I(X;T) because ∫ p(t,x) log p(x|t) dt dx = −H(X|T). Along the Markov chain X→T→T̂, the data processing inequality gives I(T̂;X) ≤ I(T;X), the opposite direction. Therefore Eq. (9b) is generally false, and the subsequent variational lower bound in Eqs. (10)–(11) is at best a bound on I(X;T), not on the distortion term I(T̂;X) in the GenIB objective (4)–(5). Consequently, the training loss in Eqs. (12)–(13) does not follow from the stated constrained optimization; it reduces to a β-VAE-style reconstruction-plus-KL objective. This invalidates the theoretical foundation of GenIB-based tokenization as presented.
[Algorithm 1 and Eq. (13)] The loss function in Eq. (13) is evaluated on the pre-channel stochastic token t (and the deterministic mean μ), i.e., CE(Fβ(t), x) and CE(Fβ(μ), x). In contrast, Algorithm 1 (lines 7–9) transmits z_m over the wireless channel, obtains t̂_m, and reconstructs x̂ = F_β^{-1}(t̂_m); however, line 10 then calculates the distortion 'based on (14)', which uses the sampled pre-channel tokens t_k, not t̂_m. Thus the channel noise and fading never appear in the gradient of the tokenizer or de-tokenizer. Unless the loss is actually computed on t̂ (which would contradict Eq. (13) as written), the training is not end-to-end through the wireless channel, undermining the claim that UniToCom learns representations robust to dynamic channel conditions. The manuscript must clarify which quantity enters the loss and, if the channel is intended to be part of training, revise Eq. (13) and Algorithm 1 accordingly.
[Section V, Simulation Results] The empirical claims in Section V are not supported by the reporting. No numerical results (e.g., tables of accuracy, FID, or WER with error bars) are given; the figures in the submitted text have garbled axis labels and legends, making quantitative assessment impossible. The assertion that UniToCom 'approaches the performance upper bound' cannot be verified without numbers or a clear description of the upper bound. At minimum, the authors should provide a table of mean and standard deviation for each metric and SNR, and specify the channel model, SNR range, and baseline configurations.

minor comments (5)

[Section III-A, Eq. (6a)] The notation I(T, X) is used in Eq. (6a), while I(X;T) is used elsewhere; please standardize to I(X;T) throughout.
[Section III-B, variance definition] The text states pα(t|x) = N(t|μ, σ) with σ a fixed variance, but then writes 'σ ∼ N(0, Cσ)', making σ both a fixed hyperparameter and a random variable. This is confusing and should be clarified (e.g., use σ² as a fixed scalar and a separate variable for the sampled noise).
[References] References [3] and [5] are both cited as arXiv:2505.10946; one of them likely has a different identifier and should be corrected.
[Section II-B, Eq. (3)] The de-tokenizer F^{-1}_m in Eq. (3) is later denoted F_β^{-1} in Section III-A; please make the notation consistent.
[Section IV-B, Eq. (18)] The noise schedule uses ω_r = ∏_{i=1}^r (1 − φ_i), but the forward process is written as q(c^r_i|c^{r-1}_i) = N(c^r_i | √(1−φ_r)c^{r-1}_i, φ_r I); the relation between the two is standard but should be stated explicitly to avoid confusion.

Circularity Check

2 steps flagged · score 6.0 of 10

GenIB derivation reverses DPI and reduces the loss to the standard beta-VAE objective; the full UniToCom system retains independent empirical content.

other [Section III-B, Eqs. (9a)-(9b)]
"Next, the distortion term I( ˆT , X) is considered as I( ˆT , X) =−H(X| ˆT ) +H(X) (9a) ⩾ R p(t, x) logp(x|t)dtdx + H(X), (9b) where (9b) holds due to the data processing inequality."

Along the paper's own chain X→T→T̂, the data processing inequality gives I(T̂;X) ≤ I(T;X). The right side of (9b) equals −H(X|T)+H(X) = I(T;X), so (9b) asserts the opposite direction and is not a consequence of DPI. Because this quantity is then used as the distortion term in (5), the 'distortion' and 'rate' terms become the same mutual information up to the variational approximation; the optimization no longer represents a rate–distortion tradeoff between compression and receiver-side informativeness. The derivation therefore collapses into a single information quantity rather than establishing the proposed GenIB tradeoff.
renaming known result [Section III-B, Eq. (12)]
"By integrating (8) and (11b) into (5), the objective function of tokenization is formalized as LGenIB(α, β) = ξDKL (pα(t|x)||N (t|0, I)) − R p(t, x) log qβ(x|t)dtdx. (12)"

This is exactly the standard β-VAE objective: a KL divergence weighted by ξ plus a negative reconstruction log-likelihood, and the paper itself notes that the conditional entropy in (12) 'is commonly solved by the mean-square-error and cross-entropy loss.' Presenting this as the newly proposed GenIB principle renames a known variational objective rather than deriving a new generative information bottleneck. The later σ-GenIB loss in Eq. (13) adds a deterministic-path term, but the stated GenIB principle itself reduces by construction to a pre-existing objective, so the theoretical derivation is not independent of prior work.

full rationale

The self-citation [9] is minor and not load-bearing: it motivates the need for a generative IB but the derivation does not depend on it. The substantive issue is the GenIB derivation itself. Equation (9b) reverses the data processing inequality, and the resulting Eq. (12) is the standard β-VAE loss by construction. Thus the paper's central first-principles claim, the GenIB principle, is a repackaging of a known result rather than a derived new tradeoff; this is a partial circularity of the theoretical derivation. However, the full UniToCom system contains independent components — the fixed-variance deterministic/stochastic training in σ-GenIB, the causal Transformer MLLM, and end-to-end channel-aware training — and the empirical comparisons against VAE and SemCom baselines do not rely on the faulty bound. No fitted-parameter-called-prediction pattern is present; ξ and λ are hyperparameters. The score reflects that the GenIB derivation reduces by construction, while the broader system retains independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central design rests on one variational bound that is applied with reversed inequality, plus several unspecified hyperparameters and a deployed 1.4B MLLM assumed available at the receiver. No independent physical entities are postulated. The simulations also rely on fixed channel and pretrained codec modules from [8].

free parameters (5)

ξ (GenIB rate weight)
Lagrange multiplier balancing compression and reconstruction in LGenIB, Eq. (5) and Eq. (13); value not reported in the manuscript.
λ (deterministic-stochastic weight)
Weights deterministic CE versus stochastic CE in Eq. (13); value not reported.
Fixed variance σ in pα(t|x)
The paper says σ is fixed, then writes σ ∼ N(0,Cσ), leaving the value and role ambiguous; needed for the σ-GenIB loss.
Cσ (variance hyperparameter)
Hyperparameter controlling the stochastic component; not specified.
Image token length Sv = 32
Optimal token length selected by scanning {16,32,64,128} in Fig. 3; the central claim of moderate compression depends on this choice.

assumptions (4)

standard math Data processing inequality holds for the chain X→T→T̂
Invoked in Eq. (9b), but applied in the wrong direction; the derivation relies on the erroneous reversed inequality.
domain assumption Gaussian latent prior q(t)=N(0,I) induces useful sparsity
Used in Eq. (8) to close the rate term; no evidence that this prior is appropriate for multimodal token transmission.
domain assumption The scalar Rayleigh fading channel model h_m z_m + n_m captures wireless dynamics
Used in Eq. (2) and simulations; does not include interference, mobility, or multi-path beyond a single scalar coefficient.
ad hoc to paper A pretrained 1.4B causal Transformer MLLM can be deployed at the receiver and integrates with the tokenizer
The receiver is assumed to have sufficient compute; no training of the MLLM with the channel is described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Token Communication in the Era of Large Models: An Information Bottleneck-Based Approach." pith.science (2026). https://pith.science/paper/KXIDSPGX

@misc{pith2026250701728,
  author       = {Pith},
  title        = {Pith review of: Token Communication in the Era of Large Models: An Information Bottleneck-Based Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXIDSPGX}},
  note         = {Machine review of arXiv:2507.01728}
}

abstract

This letter proposes UniToCom, a unified token communication paradigm that treats tokens as the fundamental units for both processing and wireless transmission. Specifically, to enable efficient token representations, we propose a generative information bottleneck (GenIB) principle, which facilitates the learning of tokens that preserve essential information while supporting reliable generation across multiple modalities. By doing this, GenIB-based tokenization is conducive to improving the communication efficiency and reducing computational complexity. Additionally, we develop $\sigma$-GenIB to address the challenges of variance collapse in autoregressive modeling, maintaining representational diversity and stability. Moreover, we employ a causal Transformer-based multimodal large language model (MLLM) at the receiver to unify the processing of both discrete and continuous tokens under the next-token prediction paradigm. Simulation results validate the effectiveness and superiority of the proposed UniToCom compared to baselines under dynamic channel conditions. By integrating token processing with MLLMs, UniToCom enables scalable and generalizable communication in favor of multimodal understanding and generation, providing a potential solution for next-generation intelligent communications.

Figures

Figures reproduced from arXiv: 2507.01728 by the authors.

**Figure 1.** The proposed UniToCom paradigm: Processing and transmitting tokens for multimodal understanding and generation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Model performance versus SNR for various tasks under Rayleigh channels. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Performance comparisons of image token length. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Networking-Aware Energy Efficiency in Agentic AI Inference: A Survey
eess.SY 2026-04 unverdicted novelty 4.0 of 10

The paper surveys energy efficiency strategies for Agentic AI inference by proposing a new accounting framework and taxonomy that spans model simplification, computation control, input optimization, and cross-layer co...

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages · cited by 1 Pith paper

[1]

Multi-Task Semantic Communications via Large Models

W. Ni, Z. Qin, H. Sun, X. Tao, and Z. Han, “Multi-task semantic communications via large models,” arxiv:2503.22064

work page Pith review arXiv
[2]

Large AI model empowered multimodal semantic communications,

F. Jiang et al. , “Large AI model empowered multimodal semantic communications,” IEEE Commun. Mag. , vol. 63, no. 1, pp. 76-82, Jan. 2025

work page 2025
[4]

Intellicise wireless networks from semantic commu- nications: A survey, research issues, and challenges,

P. Zhang et al. , “Intellicise wireless networks from semantic commu- nications: A survey, research issues, and challenges,” IEEE Commun. Surveys Tuts., vol. 27, no. 3, pp. 2051-2084, Jun. 2025

work page 2025
[5]

Next token prediction towards multimodal intelligence: A comprehensive survey,

L. Chen et al., “Next token prediction towards multimodal intelligence: A comprehensive survey,” arxiv:2505.10946

arXiv
[6]

The information bottleneck problem and its applications in machine learning,

Z. Goldfeld and Y . Polyanskiy, “The information bottleneck problem and its applications in machine learning,” IEEE J. Sel. Areas Inf. Theory , vol. 1, no. 1, pp. 19-38, May 2020

work page 2020
[7]

Transfusion: Predict the next token and diffuse images with one multi-modal model,

C. Zhou et al. , “Transfusion: Predict the next token and diffuse images with one multi-modal model,” arxiv:2408.11039

arXiv
[8]

SwinJSCC: Taming swin transformer for deep joint source-channel coding,

K. Yang et al. , “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
[9]

Federated semantic learning driven by information bottleneck for task-oriented communications,

H. Wei et al. , “Federated semantic learning driven by information bottleneck for task-oriented communications,” IEEE Commun. Lett., vol. 27, no. 10, pp. 2652-2656, Oct. 2023

work page 2023

Show all 12 references

[10]

Deep variational information bottleneck,

A. A. Alemi, I. Fischer, J. V . Dillon, and K. Murphy, “Deep variational information bottleneck,” in Proc. Int. Conf. Learn. Represent. , Toulon, France, Apr. 2017, pp. 1-19

2017
[11]

Task-oriented multi-user semantic communications,

H. Xie, Z. Qin, X. Tao, and K. B. Letaief, “Task-oriented multi-user semantic communications,” IEEE J. Sel. Areas Commun. , vol. 40, no. 9, pp. 2584-2597, Sept. 2022

2022
[12]

Gener- ative semantic communications with foundation models: Perception- error analysis and semantic-aware power allocation,

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

2025
[13]

Deep learning enabled semantic communications with speech recognition and synthesis,

Z. Weng et al. , “Deep learning enabled semantic communications with speech recognition and synthesis,” IEEE Trans. Wireless Commun. , vol. 22, no. 9, pp. 6227-6240, Sept. 2023

2023

Pith tools

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

[1] [1]

Multi-Task Semantic Communications via Large Models

W. Ni, Z. Qin, H. Sun, X. Tao, and Z. Han, “Multi-task semantic communications via large models,” arxiv:2503.22064

work page Pith review arXiv

[2] [2]

Large AI model empowered multimodal semantic communications,

F. Jiang et al. , “Large AI model empowered multimodal semantic communications,” IEEE Commun. Mag. , vol. 63, no. 1, pp. 76-82, Jan. 2025

work page 2025

[3] [4]

Intellicise wireless networks from semantic commu- nications: A survey, research issues, and challenges,

P. Zhang et al. , “Intellicise wireless networks from semantic commu- nications: A survey, research issues, and challenges,” IEEE Commun. Surveys Tuts., vol. 27, no. 3, pp. 2051-2084, Jun. 2025

work page 2025

[4] [5]

Next token prediction towards multimodal intelligence: A comprehensive survey,

L. Chen et al., “Next token prediction towards multimodal intelligence: A comprehensive survey,” arxiv:2505.10946

arXiv

[5] [6]

The information bottleneck problem and its applications in machine learning,

Z. Goldfeld and Y . Polyanskiy, “The information bottleneck problem and its applications in machine learning,” IEEE J. Sel. Areas Inf. Theory , vol. 1, no. 1, pp. 19-38, May 2020

work page 2020

[6] [7]

Transfusion: Predict the next token and diffuse images with one multi-modal model,

C. Zhou et al. , “Transfusion: Predict the next token and diffuse images with one multi-modal model,” arxiv:2408.11039

arXiv

[7] [8]

SwinJSCC: Taming swin transformer for deep joint source-channel coding,

K. Yang et al. , “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

[8] [9]

Federated semantic learning driven by information bottleneck for task-oriented communications,

H. Wei et al. , “Federated semantic learning driven by information bottleneck for task-oriented communications,” IEEE Commun. Lett., vol. 27, no. 10, pp. 2652-2656, Oct. 2023

work page 2023